Student Projects
Undergraduate students in the third year of the Final Honour School of Computer Science may undertake a project. Fourth years of the Final Honour School of Computer Science, and students for the MSc in Computer Science are required to undertake a project. Mathematics & Computer Science undergraduates are required to undertake a Computer Science project or a Mathematics dissertation in their fourth year. Computer Science & Philosophy undergraduates may choose to undertake a Computer Science project or a Philosophy thesis in their fourth year.
This is your chance to work on your own project, something that interests and inspires you. We have put together a brief guidance document on how to do it; please click on the link below. On this site, you will also find a sortable and searchable list of projects and potential supervisors. Undergraduate students are welcome to choose one of the projects on the list, or approach a potential supervisor and negotiate your own topic. MSc students should discuss with potential supervisors and get agreement.
You will be expected to make arrangements with potential supervisors between weeks 5 and 7 in Hilary Term, for project work for the following year. Please don't contact potential supervisors outside this window.
Limited equipment is available for use in projects, namely LEAP equipment. If you have any requests, please contact the Head of Academic Administration.
Important Deadlines are:
 All:
 Monday of Week 7, Hilary Term : your online project registration survey must be completed. You will be sent a link.
 MSc students:
 Monday of Week 1, Trinity Term : submission deadline for your project proposal.
 3rd/4th year students:
 12pm on Monday of Week 4 of Trinity Term : Submission deadline.
 MSc students:
 Tuesday of Week 5 (minus 5), Michaelmas Term : Submission deadline
Project writing handbook
Sample projects
3rd year
4th year
List of projects
Suitable for
Research themes
Supervisors
Project  Supervisors  Themes  B  C  MSc  Description  

Modelling Delayed Labels in Online Continual Learning  MSc  Cosupervised by Phillip Torr and Adel Bibi from the Department of Engineering Science. Online continual learning is the problem of predicting every sample in the stream while simultaneously learning from it. That is to say, the stream first presents data to be predicted and then the stream reveals the labels for the model to train on. However, in most realcase scenarios, labelling is an expensive laborious and timeconsuming procedure. Thereof, we seek to study the sensitivity of existing continual learning algorithms when labels of images at step t are only provided at step t + k. This setting poses two challenges. (1) Learning from unlabelled data. (2) Modelling the delayed labels. To that end, we are interested in proposing new algorithms that per time step t can perform selfsupervision continually while jointly training on the labelled data revealed from step tk. 

Aggregation of Photovoltaic Panels  Alessandro Abate  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  The increased relevance of renewable energy sources has modified the behaviour of the electrical grid. Some renewable energy sources affect the network in a distributed manner: whilst each unit has little influence, a large population can have a significant impact on the global network, particularly in the case of synchronised behaviour. This work investigates the behaviour of a large, heterogeneous population of photovoltaic panels connected to the grid. We employ Markov models to represent the aggregated behaviour of the population, while the rest of the network (and its associated consumption) is modelled as a single equivalent generator, accounting for both inertia and frequency regulation. Analysis and simulations of the model show that it is a realistic abstraction, and quantitatively indicate that heterogeneity is necessary to enable the overall network to function in safe conditions and to avoid load shedding. This project will provide extensions of this recent research. In collaboration with an industrial partner. Prerequisites: ComputerAided Formal Verification, Probabilistic Model Checking 

Analysis and verification of stochastic hybrid systems  Alessandro Abate  Automated Verification  C  MSc  Stochastic Hybrid Systems (SHS) are dynamical models that are employed to characterize the probabilistic evolution of systems with interleaved and interacting continuous and discrete components. Formal analysis, verification, and optimal control of SHS models represent relevant goals because of their theoretical generality and for their applicability to a wealth of studies in the Sciences and in Engineering. In a number of practical instances the presence of a discrete number of continuously operating modes (e.g., in faulttolerant industrial systems), the effect of uncertainty (e.g., in safetycritical airtraffic systems), or both occurrences (e.g., in models of biological entities) advocate the use of a mathematical framework, such as that of SHS, which is structurally predisposed to model such heterogeneous systems. In this project, we plan to investigate and develop new analysis and verification techniques (e.g., based on abstractions) that are directly applicable to general SHS models, while being computationally scalable. Courses: ComputerAided Formal Verification, Probabilistic Model Checking, Probability and Computing, Automata Logic and Games Prerequisites: Familiarity with stochastic processes and formal verification 

Automated verification of complex systems in the energy sector  Alessandro Abate  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  Smart microgrids are smallscale versions of centralized electricity systems, which locally generate, distribute, and regulate the flow of electricity to consumers. Among other advantages, microgrids have shown positive effects over the reliability of distribution networks. These systems present heterogeneity and complexity coming from 1. local and volatile renewables generation, and 2. the presence of nonlinear dynamics both over continuous and discrete variables. These factors calls for the development of proper quantitative models. This framework provides the opportunity of employing formal methods to verify properties of the microgrid. The goal of the project is in particular to focus on the energy production via renewables, such as photovoltaic panels. The project can benefit form a paid visit/internship to industrial partners. Courses: ComputerAided Formal Verification, Probabilistic Model Checking, Probability and Computing, Automata Logic and Games Prerequisites: Familiarity with stochastic processes and formal verification, whereas no specific knowledge of smart grids is needed. 

Bayesian Reinforcement Learning: Robustness and Safe Training  Alessandro Abate  Artificial Intelligence and Machine Learning, Automated Verification  C  In this project we shall build on recent work on ``Safe Learning'' [2], which frames classical RL algorithms to synthesise policies that abide by complex tasks or objectives, whilst training safely (that is, without violating given safety requirements). Tasks/objectives for RLbased synthesis can be goals expressed as logical formulae, and thus be richer than standard rewardbased goals. We plan to frame recent work by OXCAV [2] in the context of Bayesian RL, as well as to leverage modern robustness results, as in [3]. We shall pursue both modelbased and free approaches. [2] M. Hasanbeig, A. Abate and D. Kroening, ``Cautious Reinforcement Learning with Logical Constraints,'' AAMAS20, pp. 483491, 2020. [3] B. Recht, ``A Tour of Reinforcement Learning: The View from Continuous Control,'' Annual Reviews in Control, Vol. 2, 2019.  
Formal verification of a software tool for physical and digital components  Alessandro Abate  Automated Verification  MSc  We are interested in working with existing commercial simulation software that is targeted around the modelling and analysis of physical, multidomain systems. It further encompasses the integration with related software tools, as well as the interfacing with devices and the generation of code. We are interested in enriching the software with formal verification features, envisioning the extension of the tool towards capabilities that might enable the user to raise formal assertions or guarantees on models properties, or to synthesise correctbydesign architectures. Within this longterm plan, this project shall target the formal generation of faults warnings, namely of messages to the user that are related to ``bad (dynamical) behaviours'' or to unwanted ``modelling errors''. The student will be engaged in developing algorithmic solutions towards this goal, while reframing them within a formal and general approach. The project is interdisciplinary in dealing with hybrid models involving digital and physical quantities, and in connecting the use of formal verification techniques from the computer sciences with more classical analytical tools from control engineering Courses: ComputerAided Formal Verification, Software Verification. Prerequisites: Knowledge of basic formal verification. 

Model learning and verification  Alessandro Abate  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  This project will explore connections of techniques from machine learning with successful approaches from formal verification. The project has two sides: a theoretical one, and a more practical one: it will be up to the student to emphasise either of the two sides depending on his/her background and/of interests. The theoretical part will develop existing research, for instance in one of the following two interdisciplinary domain pairs: learning & repair, or reachability analysis & Bayesian inference. On the other hand, a more practical project will apply the above theoretical connections on a simple models setup in the area of robotics and autonomy. Courses: ComputerAided Formal Verification, Probabilistic Model Checking, Machine Learning 

Precise simulations and analysis of aggregated probabilistic models  Alessandro Abate  Automated Verification  C  MSc  This project shall investigate a rich research line, recently pursued by a few within the Department of CS, looking at the development of quantitative abstractions of Markovian models. Abstractions come in the form of lumped, aggregated models, which are beneficial in being easier to simulate or to analyse. Key to the novelty of this work, the proposed abstractions are quantitative in that precise error bounds with the original model can be established. As such, whatever can be shown over the abstract model, can be as well formally discussed over the original one. This project, grounded on existing literature, will pursue (depending on the student's interests) extensions of this recent work, or its implementation as a software tool. Courses: ComputerAided Formal Verification, Probabilistic Model Checking, Machine Learning 

Reinforcement Learning for Space Operations  Alessandro Abate, Licio Romao  Artificial Intelligence and Machine Learning, Automated Verification  C  Oxcav has an ongoing collaboration with the European Space Agency (ESA) that involves applying Reinforcement Learning (RL) algorithms to a satellite called OPSSAT, which has been launched in 2019 and is a flying laboratory that allows ESA’s partners to test and validate new techniques (more information can be found at https://www.esa.int/Enabling_Support/Operations/OPSSAT). This project aims at designing controllers that will be used to perform nadir pointing and suntracking of OPSSAT, while meeting some specifications (e.g., admissible nadir pointing errors). The focus will be on datadriven methods that leverage available sensors (gyroscopes, GPS, fine sun sensor, magnetometer) and actuators data using a RL architecture to come up with a safe policy that can yield an adequate performance. The main tasks of the project will consist in (1) exploring an ESA platform called MUST to collect all the necessary data and (2) implementing a RL scheme that will be later deployed in the satellite. Throughout the project you will have the opportunity to work with stateoftheart datadriven techniques that have been developed at Oxcav, under the supervision of Prof. Alessandro Abate and Dr. Licio Romao.  
Safe Reinforcement Learning  Alessandro Abate  Automated Verification  C  MSc  Reinforcement Learning (RL) is a known architecture for synthesising policies for Markov Decision Processes (MDP). We work on extending this paradigm to the synthesis of ‘safe policies’, or more general of policies such that a linear time property is satisfied. We convert the property into an automaton, then construct a product MDP between the automaton and the original MDP. A reward function is then assigned to the states of the product automaton, according to accepting conditions of the automaton. With this reward function, RL synthesises a policy that satisfies the property: as such, the policy synthesis procedure is `constrained' by the given specification. Additionally, we show that the RL procedure sets up an online value iteration method to calculate the maximum probability of satisfying the given property, at any given state of the MDP. We evaluate the performance of the algorithm on numerous numerical examples. This project will provide extensions of these novel and recent results. Prerequisites: ComputerAided Formal Verification, Probabilistic Model Checking, Machine Learning 

Safety verification for space dynamics via neuralbased control barrier functions  Alessandro Abate  Automated Verification  C  Barrier functions are Lyapunovlike functions that serve as certificates for the safety verification of dynamical and control models. The OXCAV group has recently worked on the automated and sound synthesis of barrier functions structured as neural nets, with an approach that uses SAT modulo theory. In this project, we shall pursue two objectives: 1. Apply recent results [1] on sound and automated synthesis of barrier certificates on models that are pertinent to the Space domain, such as models for attitude dynamics. Airbus will support this part. 2. Develop new results that extend theory and algorithms to models encompassing uncertainty, such as probabilistic models or models that are adaptive to sensed data. Airbus will support this effort providing environments and models for experiments and testing. [1] A. Abate, D. Ahmed and A. Peruffo, ``Automated Formal Synthesis of Neural Barrier Certificates for Dynamical Models,'' TACAS, To appear, 2021.  
Software development for abstractions of stochastic hybrid systems  Alessandro Abate  Automated Verification  C  MSc  Stochastic hybrid systems (SHS) are dynamical models for the interaction of continuous and discrete states. The probabilistic evolution of continuous and discrete parts of the system are coupled, which makes analysis and verification of such systems compelling. Among specifications of SHS, probabilistic invariance and reachavoid have received quite some attention recently. Numerical methods have been developed to compute these two specifications. These methods are mainly based on the state space partitioning and abstraction of SHS by Markov chains, which are optimal in the sense of reduction in abstraction error with minimum number of Markov states. The goal of the project is to combine codes have been developed for these methods. The student should also design a nice user interface (for the choice of dynamical equations, parameters, and methods, etc.). Courses: Probabilistic Model Checking, Probability and Computing, Numerical Solution of Differential Equations Prerequisites: Some familiarity with stochastic processes, working knowledge of MATLAB and C 

AI Vulnerability Modelling  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  This project would seek to study AI systems to identify weak points that might make implementations open to compromise and attack, and to develop corresponding threat models. The approach might involve designing test suites using attack graphs and validation in a laboratory setting.  
AugmentedReality Personal Security Solutions  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  This project seeks to explore the application of augmentedreality technologies to developing more usable personalsecurity solutions. There is a growing body of work exploring how interactive augmentedreality technologies can help users to maintain awareness of potential threats and thus protect themselves against cyberattacks. The project will build on existing work to develop and validate the effectiveness of new augmentedreality tools. 

Chatbot Attack and Vulnerability Models  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  As chatbots are increasingly widely used in a growing range of applications, understanding the potential for cyberattacks to compromise the integrity and confidentiality of the data they output is critical. This project would seek to study chatbot systems to identify weak points that might make implementations open to compromise and attack, and to develop corresponding threat models. The approach might involve designing test suites using attack graphs and validation in a laboratory setting.  
Deep Learning Models to Support Computer Network Defence  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  This project would explore the application of deeplearning models such as Large Language Models (LLM) to computer network defence. Deep learning has the potential to improve processes at various defensive stages including identification of assets and risk analysis; detection and mitigation of attacks; and automated response. The project scope is broad enough to allow the student to identify an area of interest.  
Designing Cybersecurity Test Suites for Generative AI Systems  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  As generative AI systems are increasingly widely used in a growing range of applications, understanding the potential for cyberattacks to compromise the integrity and confidentiality of the data they output is critical. This project would aim to develop attack graphs for generative AI systems, and based on this design cybersecurity test suites that facilitate testing the security of generate AI systems against a range of potential attacks. The project might involve using these test suites to test the security of a range of generative AI implementations in a laboratory setting.  
Generating Realistic Cybersecurity Datasets and Testbeds  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  A common challenge in the cybersecurity research community is the limited availability of cybersecurity data: while there are some real data captures made available publicly or for research purposes, challenges of confidentiality and practicality mean the resources are limited. A particular example is the availability of network data from various environments: the networks of organisations; Industrial InternetofThings (IIoT) environments; and Blockchain networks. This project seeks to develop highquality synthetic cybersecurity datasets that would benefit the cybersecurity research community. The project would begin with background research to identify the characteristics of a realistic dataset. The approach might then involve creating a cybersecurity testbed composed of a synthetic network and set of attack test suites, and capturing traffic.  
Insider threat detection  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  Organisations are experiencing an evergrowing concern of how to identify and defend against insider threats. Those who have authorised access to sensitive organisational data are placed in a position of power that could well be abused and could cause significant damage to an organisation. This could range from financial theft and intellectual property theft, through to the destruction of property and business reputation. This focus of this project is on the detection and identification of insider threats to organisations. The project builds on the group’s previous research in this area.  
LowOrbit Space Cybersecurity  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  This project would seek to study loworbit space systems to identify weak points that might make implementations open to compromise and attack, and to develop corresponding threat models. The approach might involve designing test suites using attack graphs and validation in a laboratory setting.  
Operational Security Tools for Users with Limited Cybersecurity Knowledge  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  As digitisation increases globally, there is a growing range of different types of user required to have an understanding of the operational cybersecurity of digital systems – i.e., the ability to detect and respond to potential cyber attacks. This includes not only individual users of smartphones and computers, but also personnel in organisations from the growing range of sectors becoming digitised (e.g., transport, marine, energy, healthcare). New challenges are emerging in terms of how to enable parties to monitor the necessary information and make informed security decisions. This project seeks to identify the requirements of users in this area, and develop operational cybersecurity tools developed to this need. The project scope is broad, and the student may select particular sectors of interest, or particular approaches of interest such as cybersecurity visualizations or automated incidentresponse tools.  
Ransomware detection  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  Ransomware attacks are increasing rapidly every year. While signaturebased malware detection methods work well for detecting and stopping known malware, attackers can bypass the detection using obfuscation techniques or zeroday attacks. There is therefore a need for better detection mechanisms that are able to predict new forms of malware and react against them. This project aims at exploring malware detection to develop a better understanding of the differences that make malware different from normal processes or files. It will further seek to implement a machine learning (ML) tool that would help in detecting malicious behaviour efficiently, so that malware infection and propagation can be stopped. The ML classifiers used will depend on the malware family explored as well as data available. 

SystemicRisk Modelling  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  A number of features of the digital environment create the potential for systemic risk, i.e., the possibility that a single event or development might trigger widespread failures and negative effects spanning multiple organisations, sectors or nations. These features include the interconnectivity of digital systems, homogeneity of devices meaning vulnerabilities may affect a large proportion of systems, and high dependency of organisations on service providers. This project would aim to develop models that allow analysis and quantification of the systemic risk that a population may face. The approaches used may include agentbased simulations and/or the development of harm trees.  
Visualising Large Cybersecurity Datasets  Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  A key challenge for cybersecurity operations in organisations is the sheer scale of the data generated within organisations as they increasingly digitise their operations. The scale of data makes is challenging for personnel responsible for monitoring security to maintain a clear situational awareness of network activity, and detect, identify and respond to potential attacks. This project focuses on identifying and implementing approaches to visualising large cybersecurity datasets in a way that is comprehensible and digestible for users. The project might include requirements gathering (e.g., through interviews) to identify users’ needs; implementation and validation of visualization approaches in a laboratory setting; or testing of the developed visualizations with end users to assess utility.  
Sonification for detecting cyberattacks  Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  In the face of increasingly frequent, sophisticated and varied cyberattacks, organisations must continuously adapt and improve their network defences. There is a need for effective tools to help security practitioners to engage with and understand the data communicated over the network, and the outputs of automated attackdetection methods. Visual (textbased and graphical) presentations of data are usually used for this. Over the last few years, the utility of sonification (the representation of data as sound) for networksecurity monitoring has been theorised and explored experimentally. This project seeks to build on prior research in our research group and externally, in which the effectiveness of sonification at representing a limited range of attack types has been experimented with. The aim of the project is to expand on this experimentation by assessing and comparing the effectiveness of various sonification designs at representing a wider range of attack types. This will involve identifying the key indicators present in network data for each attack type, exploring how the sonification design can best represent these indicators, and experimenting with the effectiveness of the resulting sonification approaches. Attack types experimented with could include, for example, various malwares and ransomwares, advanced persistent threats, and various types of flooding attack (this could use both real and synthetic attack datasets).


Threat Models for Blockchains  Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  This project would seek to study Blockchainbased systems to identify weak points that might make implementations open to compromise and attack, and to develop corresponding threat models. The project might focus on the general form of Blockchains, or focus on a particular implementations or consensus protocols. A number of different approaches might be considered, for example: formal modelling of protocols used to establish consensus or resolve conflicts; design of test suites using attack graphs and validation in a laboratory setting.  
Complex Object Querying and Data Science  Michael Benedikt  Data, Knowledge and Action  MSc  "We will look at query languages for transforming nested collections (collections that might contain collections). Such languages can be useful for preparing large scale feature data for machine learning algorithms. We have a basic implementation of such a language that we implement on top of the bigdata framework Spark. The goal of the project is to extend the language with iteration. One goal will be to look at how to adapt processing techniques for nested data to support iteration. Another, closer to application is to utilize iteration to support additional steps of a data science pipeline, such as sampling. "  
Decision procedures for arithmetic with powers  Michael Benedikt  Data, Knowledge and Action  C  MSc  The project will look at implication and satisfaction problems involving formulas that involve additive arithmetic, inequalities, and some form of exponentiations with a constant base. For example, systems of inequalities with constant multiples of variables and constant multiple of expressions like 2^x. We will also look at first order logic built up from these inequalities. In the absence of exponentiatial terms like 2^x, the theory is known as Presburger Arithmetic, and has been heavily investigated in both theory and practice. In the presence of expenontial terms to a fixed base, it is known to be decidable, from work of Semenov in the late 70's and 80's. Recent work has shown fragments where the complexity is not too terrible, and in the process some new algorithms have been proposed. The goal of the project is to implement decision procedures for expressions of this form. A theoreticallyoriented student can also work on complexity analysis. A prerequisite for the project is good background in logic  Logic and Proof and preferably one of CAFV or KRR.  
Graph Machine Learning with Neo4j  Michael Benedikt  Data, Knowledge and Action  MSc  Cosupervisor Neo4J (industrial partner) We have discussed a set of projects concerning data science on graphs with Brian Shi (an Oxford alumni) of the Neo4j data science team, including (but not limited to):  understanding and evaluating different initialization regimes in node embeddings,  explainability and interpretability for graph ML, design of online grap ML algorithms/models where nodes and edges are addfed/modified in near realtimel A student would be working with Michael Benedikt and the Neo4j team: the project will be researchfocused and any software produced would be in the public domain. The balance of experiment and theory could be tuned to the student's interests. The main prerequisite would be a very good knowledge of graph ML, at the level of Oxford's GRL course. 

Interpolation  Michael Benedikt  Data, Knowledge and Action  C  MSc  Let F1 and F2 be sentences (in firstorder logic, say) such that F1 entails F2: that is, any model of F1 is also a model of F2. An interpolant is a sentence G such that F1 entails G, and G entails F2, but G only uses relations and functions that appear in *both* F1 and F2. The goal in this project is to explore and implement procedures for constructing interpolants, particularly for certain decidable fragments of firstorder logic. It turns out that finding interpolants like this has applications in some database query rewriting problems. Prerequisites: Logic and Proof (or equivalent) 

Optimization of Web Query Plans  Michael Benedikt  Data, Knowledge and Action  MSc  This project will look at how to find the best plan for a query, given a collection of data sources with access restrictions. We will look at logicbased methods for analyzing query plans, taking into account integrity constraints that may exist on the data. 

Optimized reasoning with guarded logics  Michael Benedikt  Data, Knowledge and Action  C  MSc  "Inference in firstorder logic is undecidable, but a number of logics have appeared in the last decade that achieve decidability. Many of them are guarded logics, which include the guarded fragment of firstorder logic. Although the decidability has been known for some decades, no serious implementation has emerged. Recently we have developed new algorithms for deciding some guarded logics, based on resolution, which are more promising from the perspective of implementation. The project will pursue this both in theory and experimentally." Prerequisites A knowledge of firstorder logic, e.g. from the Foundations of CS or Knowledge Representation and Reasoning courses, would be important.  
Bypassing PostFiltering Guardrails in Large Language Models.  Adel Bibi, Philip Torre, Atılım Güneş Baydin  Artificial Intelligence and Machine Learning  MSc  This task aims to assess the reliability of postfiltering methods used to prevent language models from generating specific content. These methods involve monitoring generated content for certain keywords and pausing the model when detected. The project is on investigating whether adversaries can bypass these filters, leading to the generation of undesirable content while avoiding the filtered words. This research will include the formalization of the problem and the development of efficient solutions to address this challenge. This task helps evaluate existing safety measures and explore the potential for enhancing safeguards against automated harmful content generation. It also has applications in undermining watermarks, as they rely on less common data subsets, making detection easier. 

Data Leakage in Large Language Models through Prompting  Adel Bibi, Philip Torre, Atılım Güneş Baydin  Artificial Intelligence and Machine Learning  MSc  Many online services rely on large language models finetuned with concealed system prompts detailing operational behavior. Adversaries may attempt to synthesize input to uncover these hidden prompts and manipulate the model, posing security risks. Instances of online games and competitions already exist, attempting to reveal hidden passwords in system prompts. The project is about formalizing this problem and developing systematic approaches to generating adversarial prompts that reveal vulnerabilities. This includes optimizing prompts to elicit affirmative responses indicative of data leakage. The goal is to identify minimal adversarial prompts that can bypass restrictions on prompt queries, enhancing security. 

Applications of ComplexOrder Fractional Diffusion to Physical and Biological Processes  Alfonso BuenoOrovio  Computational Biology and Health Informatics  B  C  Fractional differential equations (where standard derivatives are generalised to derivatives of noninteger, real order) have become a fundamental modelling approach for understanding and simulating the many aspects of spatial heterogeneity of complex materials and systems [1,2]. Very recently [3], we have been able to further extend such ideas to derivatives of complexorder. This project aims to further explore the capabilities of these novel complexorder fractional operators in modulating the spatiotemporal dynamics of different systems of interest. To investigate how complexorder fractional operators can advance the description of multiscale transport phenomena in physical and biological processes highly influenced by the heterogeneity of complex media. Students will investigate, by means of scientific computing and modelling and simulation, how complexorder fractional operators modulate the response of different systems of physical and biological interest. Different research options will be available, including the analysis of: (i) novel Turing patterns in physical and biological processes of morphogenesis; (ii) periodic precipitation patterns in the manufacturing of micro and nanostructures; or (iii) models of population dynamics and epidemiology. 1. Fractional diffusion models of cardiac electrical propagation: role of structural heterogeneity in dispersion of repolarization role of structural heterogeneity in dispersion of repolarization https://doi.org/10.1098/rsif.2014.0352 2. Fourier spectral methods for fractionalinspace reastion diffusion equations. https://doi.org/10.1007/s1054301404842 3. The ComplexOrder Fractional Laplacian: Stable Operators, Spectral Methods, and Applications to Heterogeneous Media. https://doi.org/10.20944/preprints202107.0569.v1  
High Throughput, High Resolution, and High FrameRate Analysis of Cellular Heart Function  Alfonso BuenoOrovio  Computational Biology and Health Informatics  B  C  MSc  Cosupervised by Christopher Toepfer Cardiomyocytes are the cells responsible for generating contractile force in the heart. They are highly adaptable and alter their function in response to many stimuli (electrical stimuli, calcium, contraction, organisation of cellular structures, metabolism, and transcriptional signatures). These processes are all interrelated and dynamic in live cardiomyocytes, but we lack the ability to visualise them simultaneously and correlate them in realtime. We have previously developed specialised software [1, 2] to automate highthroughput analysis of cardiomyocyte function. However, these are not optimised to work with large files and cause data fragmentation as they do not work in tandem. To develop novel software solutions for the analysis of cellular heart function for three channel, high throughput, resolution, and framerate imaging and analysis of beating cardiomyocytes. Imaging live cardiomyocytes during contraction and relaxation necessitates highframe rates (100s of FPS), which must be twinned with high resolutions (<70 nm per pixel), also allowing the simultaneous analysis of multiple well plates for highthroughput. In collaboration with the Oxford Department of Cardiovascular Medicine, and building on extensive datasets of highresolution, highrate fluorescent imaging of live cardiomyocytes, students taking this project will develop novel opensource software solutions integral to discovery and translational cardiovascular research. Different research options would be available, such as:
[1] CalTrack: HighThroughput Automated Calcium Transient Analysis in Cardiomyocytes. https://doi.org/10.1161/CIRCRESAHA.121.318868 [2] SarcTrack: an adaptable software tool for efficient largescale analysis of sarcomere function in hiPSCcardiomyocytes. https://doi.org/10.1161/CIRCRESAHA.118.314505 

Interpretable Cardiac Anatomy and Electrophysiology Modelling in Paediatric Patients Using Variational Mesh Autoencoders  Alfonso BuenoOrovio, Vicente Grau  Computational Biology and Health Informatics  C  MSc  Cardiac anatomy and function vary considerably across the human population with important implications for clinical diagnosis, treatment planning, and prognosis of disease. Consequently, many computerbased approaches have been developed to capture this variability for a wide range of applications, including explainable cardiac disease detection and prediction, dimensionality reduction, cardiac shape analysis, and the generation of virtual heart populations. Here, we will leverage on these technologies to further investigate connections between cardiac anatomy and electrocardiographic (ECG) signals in paediatric patients with Hypertrophic Cardiomyopathy (HCM), the most common hereditary heart disease and leading cause of sudden cardiac death in the young and competitive athletes. To investigate connections between patterns of cardiac hypertrophic and their manifestation in the ECG in HCM paediatric patients. Generation of virtual populations of HCM paediatric hearts, capturing the variability in cardiac shape and ECG signals in this group of highrisk patients. We have recently proposed novel variational mesh autoencoder (mesh VAE) methods as a novel geometric deep learning approach to model populationwide variations in cardiac anatomy [1], which enable direct processing of surface mesh representations of the cardiac anatomy in an efficient manner. These methods can also be extended to simultaneously learn the ECG signal of a given patient [2]. Exploiting cardiac Magnetic Resonance Imaging (MRI) and ECG resources from established clinical collaborators, this project will explore the quality and interpretability of the mesh VAE's latent space for the reconstruction and synthesis of multidomain cardiac signals in paediatric patients with HCM. It will also investigate the method's ability to generate realistic virtual populations of paediatric cardiac anatomies and ECG signals in terms of multiple clinical metrics in this group of patients. [1] Interpretable Cardiac Anatomy Modeling Using Variational Mesh Autoencoders. https://doi.org/10.3389/fcvm.2022.983868 [2] MultiDomain Variational Autoencoders for Combined Modelling of MRIBased Biventricular Anatomy and ECGBased Cardiac Electrophysiology. https://doi.org/10.3389/fphys.2022.886723 Prerequisites: Computational Medicine (recommended) 

Modelling and Simulation of Genetic Heart Disease  Alfonso BuenoOrovio  Computational Biology and Health Informatics  C  MSc  The goal of Precision Medicine is to provide therapies tailored to each patient. This is especially an urgent need in inherited heart disease, where current drug therapy mostly relies on symptom relief. In this regard, computational modelling and simulation constitutes a flexible platform for investigating the mechanisms underlying arrhythmic risk in these patients, and for the virtual screening of pharmacological therapy of both established and novel agents. To investigate mechanisms of arrhythmic risk and/or response to therapy using multiscale computational models (cell to wholeorgan) of inherited heart disease. Students will investigate, by means of computational modelling and simulation, how changes in structure and cellular function
modulate the risk of lifethreatening events and/or response to pharmacological therapy in patients with hypertrophic cardiomyopathy
(HCM). Different research options will be available, including: [1] Mechanisms of proarrhythmic abnormalities in [2] Improving the clinical understanding of hypertrophic [3] Electrophysiological and contractile effects of [4] Mechanism based therapies enable personalised Prerequisites: Computational Medicine (recommended) 

Frontiers in Graph Representation Learning  Ismail Ilkan Ceylan  Artificial Intelligence and Machine Learning  C  MSc  There are various projects available in the theme of ‘Graph Representation Learning’ (e.g., graph neural networks), please get in touch for the most uptodate information. There are multiple projects aiming for (i) designing new graph neural network models to overcome some wellknown existing limitations of graph neural networks (e.g., oversmoothing, oversquashing, inexpressiveness), (ii) improving our overall theoretical understanding of graph representation learning models (foundational aspects related to probability theory, graph theory, logic, as well as algorithmic aspects), (iii) improving the inductive bias of graph neural networks for machine learning over different types of graphs (e.g., directed, relational graphs), (iv) exploring tasks beyond widely studied ones, such as graph generation, from a foundational aspect (v) novel applications of graph represenation learning. It is important to identify a project which matches the students background and so the details are somewhat subtle to be included in a project description. There are various industrial and academic collaborators, depending on the specific direction. Most of the necessary background information is covered in the graph representation learning course offered in the Department in MT. 

Detection of cycles in the cryptographic protocol verifier ProVerif’s saturation procedure  Vincent Cheval  Security  B  C  MSc  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). The problem of verifying security properties (even simple secrecy) is undecidable in general, especially when considering an unlimited number of sessions. For this reason, tools such as ProVerif and Tamarin may not terminate, as they allow users to model unbounded number of sessions. In the vein of [1], many recent features have been introduced in ProVerif to help the tool terminate. However, most of these features require some manual intervention on the part of users, and it can be quite challenging to understand which feature to use depending on the protocol under study. Internally, ProVerif’s core algorithm consists of saturating of a set of Horn clauses with free resolution. In practice, most, if not all, cases of nontermination occur due to the saturation procedure entering an infinite loop. In this project, we aim to develop a procedure that can automatically detect a large class of loops and appropriately apply some of the ProVerif’s features to exit the detected loops. Objectives:
The library will be incorporated to the official release of ProVerif. Bibliography: Reading the papers is not mandatory to complete the project but it will provide a good context for the project. [1] Bruno Blanchet, Vincent Cheval, and Cortier Véronique. Proverif with lemmas, induction, fast subsumption, and much more. In Proceedings of the 43th IEEE Symposium on Security and Privacy (S&P’22). IEEE Computer Society Press, May 2022. [2] Bruno Blanchet. Modeling and Verifying Security Protocols with the Applied Pi Calculus and ProVerif. Foundations and Trends in Privacy and Security, 1(12):1135, October 2016. Prerequisites: Logic and Proof, Functional Programming 

Generation and verification of Proof Certificates for cryptographic protocols  Vincent Cheval  Security  C  MSc  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). Due to the critical aspect of these security protocols, it is crucial to guarantee that possible bugs in these automatic tools do not affect the correctness of the security proofs. In addition, even though the efficiency of these tools have significantly improved over the years, the verification of reallife protocols may require an large amount time and memory to terminate. For instance, when the verification requires more than 200GB of memory, as it was the case for verifying privacy properties of an extension of TLS 1.3, the reproducibility of the verification results is severely affected.To tackle this problem, the approach taken in this project is to develop proof certificates that can be verified by a machinechecked verifier. One of the main difficulty is to ensure that certificates are sufficiently small in practice (in order that they can be easily exchanged) while guaranteeing a relatively small verification time. The certificate generation and verification will be based on ProVerif internal procedure. Objectives:
We do not expect for the last objective to be fully completed as the implementation in F* may be extremely time consuming. However, a partial implementation will be elaborated. [1] https://www.fstarlang.org [2] Bruno Blanchet, Vincent Cheval, and Cortier Véronique. Proverif with lemmas, induction, fast subsumption, and much more. In Proceedings of the 43th IEEE Symposium on Security and Privacy (S&P’22). IEEE Computer Society Press, May 2022. Prerequisites: Logic and Proof, Functional Programming, (CAFV is a plus) 

Handling cryptographic primitives with complex algebraic properties  Vincent Cheval  Security  B  C  MSc  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). Internally, these tools express cryptographic messages using terms. The algebraic properties of the cryptographic primitives are then expressed by mean of an equational theory on terms. For example, the equation dec(enc(x,y),y) = x expresses that deciphering (“dec”) with a key “y” a cipher of a message “x” by the same key “y”, i.e. “enc(x,y)”, allows to retrieve the message “x”. Equality of terms modulo the equational theory is one of the core problems that these tools must solve. Although unification algorithms for many primitives with complexes algebraic properties, such as finite variant, XOR with homomorphism, abelian groups, etc are known, these algorithms only work when the equations do not mix primitives, in other words when the equations are not distinct. We aim to develop new efficient unification algorithm for the union of nondistinct equational theories. Proofs of correctness would be first done by hand but an implementation in F* [1] would be of interest (or Coq [2]). Objectives:
The implemented library will be incorporated in the tool ProVerif. [1] https://www.fstarlang.org [2] https://coq.inria.fr/ [3] FRANZ BAADER, KLAUS U. SCHULZ, Unification in the Union of Disjoint Equational Theories: Combining Decision Procedures, Journal of Symbolic Computation, Volume 21, Issue 2, 1996 [4] Bruno Blanchet, Martín Abadi, and Cédric Fournet. Automated Verification of Selected Equivalences for Security Protocols. Journal of Logic and Algebraic Programming, 75(1):351, FebruaryMarch 2008. Prerequisites: Logic and Proof, Functional Programming 

Multicore overhaul of the cryptographic protocol verifier ProVerif  Vincent Cheval  Security  B  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). Of the three aforementioned tools, ProVerif is generally considered to be the most efficient. It is also the only tool that does not rely on distributed or parallel computing. ProVerif’s internal procedure is based on the saturation of sets of Horn clauses which is inherently more suited to sequential computing. However, several timeconsuming algorithms used in the saturation procedure can greatly benefit from parallel computing, such as Horn clauses subsumption checking, unification and matching of terms modulo an equational theory. This project aims to transform many algorithms used in ProVerif using the multicore programming framework that is provided in OCaml 5. We can expect significant speedup as, for instance, Horn clauses subsumption checking represents usually 80 to 90% of the total execution time of ProVerif. One of the difficulties however will be to appropriately combine parallel computing algorithm with other optimization features already implemented that are specifically tailored for sequential computing (such as hash consing techniques). Objectives:
The library will be incorporated to the official release of ProVerif. Bibliography: Reading the papers is not mandatory to complete the project but it will provide a good context for the project. [1] Bruno Blanchet, Vincent Cheval, and Cortier Véronique. Proverif with lemmas, induction, fast subsumption, and much more. In Proceedings of the 43th IEEE Symposium on Security and Privacy (S&P’22). IEEE Computer Society Press, May 2022. [2] Bruno Blanchet. Modelling and Verifying Security Protocols with the Applied Pi Calculus and ProVerif. Foundations and Trends in Privacy and Security, 1(12):1135, October 2016. Prerequisites: Logic and Proof, Functional Programming, Concurrent Programming 

Reducing memory consumption of the cryptographic protocol verifier ProVerif with hash consing techniques  Vincent Cheval  Security  B  C  MSc  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). While the recent overhaul of ProVerif [1] significantly improved its efficiency, it did not really affect its memory consumption. Indeed complex reallife protocols, such as TLS, required between 10 to 100GB of memory depending on the type of security properties being proved. This problem of memory is crucial as we reach the limits of the capabilities of our computers. We aim to change the internal representations of messages within ProVerif by using hash consing techniques [2] that allow to maximize the sharing of memory. In particular, two messages semantically equal will thus be ensured to be also physically equal (they will point to the same address in the memory). It will require to rework most of the implemented algorithms operating on messages (resolution, unification, matching, subsumption,…). The internal algorithm of ProVerif being based mostly on the saturation of Horn clauses, a maximal sharing of memory within each Horn clause but also within the set of Horn clauses generated by ProVerif will allow us to considerably reduce the memory consumption of ProVerif. Maximizing the memory sharing within one clause is fairly straightforward. However, for sets of Horn clauses, the problem becomes much harder due to the presence of variables that sometimes need to be considered distinct within two Horn clauses. Objectives:
Bibliography: Reading the papers and slides in the bibliography is not mandatory to complete the project but it will provide a good context for the project. [1] Bruno Blanchet, Vincent Cheval, and Cortier Véronique. Proverif with lemmas, induction, fast subsumption, and much more. In Proceedings of the 43th IEEE Symposium on Security and Privacy (S&P’22). IEEE Computer Society Press, May 2022. [2] JeanChristophe Filliâtre, Sylvain Conchon. Typesafe modular hashconsing. ML 2006: 1219 [3] https://www3.risc.jku.at/education/courses/ss2018/unification/slides/02_Syntactic_Unification_Improved_Algorithms_handout.pdf Prerequisites: Logic and Proof, Functional Programming 

Verifying privacytype properties in a probabilistic setting  Vincent Cheval  Security  C  MSc  Security protocols aim at securing communications. They are used in various applications (e.g. establishment of secure channels over the Internet, secure messaging, electronic voting, mobile communications, etc.) Their design is known to be error prone and flaws are difficult to fix once a protocol is largely deployed. Hence a common practice is to analyze the security of a protocol using formal techniques and in particular automatic tools. This approach has lead to the development of several verification tools such as ProVerif (https://proverif.inria.fr), Tamarin (https://tamarinprover.github.io) and DeepSec (https://deepsecprover.github.io) that rely on symbolic models to express protocols and their security properties. Symbolic, automated verification has been a very successful approach for revealing vulnerabilities or proving the absence of entire families of attacks. For instance, these tools were used to verify the security of several major protocols (e.g. TLS 1.3, 5G protocols, Belenios electronic voting protocol). The DeepSec prover specializes in the verification of security properties that can be expressed as behavioural equivalence, such as anonymity, vote privacy, unlinkability, strong secrecy, etc. It currently implements a procedure for checking “trace equivalence” between processes. The paper that describes the theory behind DeepSec also describe a theoretical procedure that allows to prove two stronger behavioural equivalences, that are similarity and bissimilarity. The description of these procedures is still considered “theoretical” as they are heavily nondeterministic, hence preventing efficient implementations. One part of this project thus aims to design an effective and efficient procedure for checking bissimilairty and similarity. Currently, symbolic models are purely nondeterministic. For instance, the random generation of numbers are intuitively abstracted and they are assumed to be impossible to guess by the attacker. While this is generally sensible, it is not always possible to eliminate probabilities altogether: some protocols specifically rely on probabilistic coin flips in their specification, and a recent paper [2] has shown that giving the attacker the ability to use probabilistic choices can enable him to break behavioural equivalence. This paper provides the general probabilistic model but only describe a procedure for checking probabilistic equivalence in a restrictive case, based on the internal procedure of DeepSec. This article provides the general probabilistic model, but only describes a procedure for checking probabilistic equivalence in a restrictive case, based on DeepSec's internal procedure. The second part of this project will aim to design a (potentially theoretical) procedure for verifying probabilistic equivalence in the general setting. Currently, symbolic models are purely nondeterministic. For instance, the random generation of numbers are intuitively abstracted and they are assumed to be impossible to guess by the attacker. While this is generally sensible, it is not always possible to eliminate probabilities altogether: some protocols specifically rely on probabilistic coin flips in their specification, and a recent paper [2] has shown that giving the attacker the ability to use probabilistic choices can enable him to break behavioural equivalence. This paper provides the general probabilistic model but only describe a procedure for checking probabilistic equivalence in a restrictive case, based on the internal procedure of DeepSec. This article provides the general probabilistic model, but only describes a procedure for checking probabilistic equivalence in a restrictive case, based on DeepSec's internal procedure. The second part of this project will aim to design a (potentially theoretical) procedure for verifying probabilistic equivalence in the general setting. Objectives:
The implemented library will be incorporated in the tool DeepSec. [1] V. Cheval, S. Kremer and I. Rakotonirina. DEEPSEC: Deciding Equivalence Properties in Security Protocols  Theory and Practice. In Proceedings of the 39th IEEE Symposium on Security and Privacy (S&P'18), IEEE Computer Society Press, 2018. [2] Symbolic protocol verification with dice. Vincent Cheval‚ Raphaëlle Crubillé and Steve Kremer. In J. Comput. Secur.. Vol. 31. No. 5. Pages 501–538. 2023. Prerequisites: Logic and Proof, Functional Programming, Probabilistic Model Checking 

Detecting and Modelling Mirrors for 3D Scene Reconstruction  Ronnie Clark  Artificial Intelligence and Machine Learning, Systems  C  MSc  Elements such as mirrors and glass still present a major difficulty for simultaneous localisation and mapping (SLAM) and 3D reconstruction algorithms [1]. This project will address this challenge by developing a method for accurately identifying mirrors in a scene and accounting for their unique properties in the reconstruction process. The goal will be to improve the fidelity and accuracy of the 3D models. The work will involve aspects of computer vision, geometric modelling, and possibly ray tracing techniques. [1] https://www.thomaswhelan.ie/Whelan18siggraph.pdf Prerequisites: Suitable for those who have taken a course in machine learning or computer graphics 

Enhanced Single Image Depth Prediction using a Percentilebased Loss  Ronnie Clark  Artificial Intelligence and Machine Learning, Systems  C  MSc  Single image depth perception is a common task in computer vision which involves predicting the distance from the camera to the scene for each pixel in an image. This problem has broad applications in autonomous systems, virtual and augmented reality, as well as graphics. However, a significant challenge in single image depth prediction is the estimation of the absolute scene scale. Therefore, most depth estimation methods focus on predicting relative depth rather than absolute measurements [1]. This project aims to explore a new relative depth parameterisation based on the percentile ranking of depth values. The student will be expected to train a model using this parameterisation and evaluate its effectiveness by comparing its performance with existing methods on widelyused datasets such as NYUv2. [1] https://arxiv.org/pdf/1907.01341v3.pdf [2] https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html Prerequisites: Suitable for those who have taken a course in machine learning 

Low Rank training of Neural Fields  Ronnie Clark  Artificial Intelligence and Machine Learning, Systems  C  MSc  Neural fields [1] have emerged as a promising method for representing 3D data, utilising MultiLayer Perceptrons (MLPs) to predict the properties of a field at every point in space. Despite their potential, a significant drawback is the lengthy training process, often due to the large size of the MLPs. This project aims to explore the application of lowrank adaptation (LoRA) [2] in enhancing the training efficiency of neural fields. An essential part of this investigation will involve benchmarking the performance of lowrank training against full training across various types of fields, such as Signed Distance Functions (SDFs) or radiance fields, to evaluate the effectiveness of LoRA in this context. [1] https://neuralfields.cs.brown.edu/siggraph23.html [2] https://arxiv.org/pdf/2106.09685.pdf Prerequisites: Suitable for those who have taken a course in machine learning or computer graphics 

Unsupervised Visual Learning using SegmentMasks  Ronnie Clark  Artificial Intelligence and Machine Learning, Systems  C  MSc  Masked language modelling is the prevalent method for pretraining large language models (LLMs). This involves 'masking' words in the text and prompting the model to predict the hidden words. This approach has also been adapted in computer vision, notably in Vision Transformers (ViTs) [1], where patches of images are dropped, and the model is tasked with predicting these missing regions. However, in the vision context process isn't ideal, as patches in images don't correspond directly to words in text. Unlike words, which represent complete concepts, image patches can contain parts of various objects. This project will investigate a novel approach to vision pretraining: using segmentation masks instead of patches. A key aspect of the project will be comparing the effectiveness of segmentationbased training against the traditional patchbased method in downstream tasks, such as semantic segmentation or depth estimation. [1] https://arxiv.org/pdf/2111.06377.pdf Prerequisites: Suitable for those who have taken a course in machine learning 

Topics in Online Algorithms and LearningAugmented Algorithms  Christian Coester  Algorithms and Complexity Theory  B  C  MSc  Description: An online algorithm is an algorithm that receives its input over time, e.g. as a sequence of requests that need to be served. The algorithm must act on each request upon its arrival, without knowledge of future requests, often with the goal of minimising some cost. (For example, assigning taxis to ride requests with the goal of minimising distance traveled by taxis.) Due to the lack of information about future requests, the algorithm cannot always make optimal decisions, but for many problems there exist algorithms that provably find solutions whose cost is within a bounded factor of the optimum, regardless of the input. An emerging related field of research are learningaugmented algorithms, where the algorithm receives some (perhaps erroneous) predictions about the future as additional input. In this project, the student will work on a selected problem in the field of online algorithms or learningaugmented algorithms, with the goal of designing algorithms and theoretically analysing their performance. The project is suitable for mathematically oriented students with an interest in proving theorems about the performance of algorithms.  
REASONING ABOUT TEMPORAL KNOWLEDGE  Bernardo Cuenca Grau, Przemysław Wałęga  Data, Knowledge and Action  MSc  "Our world produces nowadays huge amounts of time stamped data, say measurements from meteorological stations, recording of online payments, GPS locations of your mobile phone, etc. To reason on top of such massive temporal datasets effectively we need to provide a wellstructured formalisation of temporal knowledge and to devise algorithms with good computational properties. This, however, is highly nontrivial; in particular logical formalisms for temporal reasoning often have high computational complexity. This project provides an opportunity to join the Knowledge Representation and Reasoning group and participate in exciting EPSRCfunded research on temporal reasoning, temporal knowledge graphs, and reasoning over streaming data. There are opportunities to engage both in theoreticallyoriented and practicallyoriented research in this area. For example, in recent months, we have been investigating the properties and applications of DatalogMTLa temporal extension of the wellknown Datalog rule language which is wellsuited for reasoning over large and frequentlychanging temporal datasets; the project could focus on analysing the theoretical properties of DatalogMTL and its fragments such as its complexity and expressiveness, or alternatively in more practical aspects such as optimisation techniques for existing reasoning algorithms. There are many avenues for research in this area and we would be more than happy to discuss possible concrete alternatives with the student(s)." The theoretical part of the project requires good understanding of logics (completing Knowledge Representation and Reasoning course could be beneficial), whereas the practical part is suited for those who have programming skills. 

Realistic Benchmarks for Explainable GNNBased Models  Bernardo Cuenca Grau, David Tena Cucala  Data, Knowledge and Action  B  C  Graph Neural Networks (GNNs) are a family of machine learning models that have been successfully applied to many realworld tasks involving graphstructured data. The predictions made by GNNs, however, are often difficult to explain because of the opaque nature of the model. This becomes especially problematic in safetycritical applications or in contexts that require compliance with principles of explainability. The Data & Knoweldge Group has recently implemented a GNNbased system where every prediction made by the model can be explained by a rule expressed in Datalog, a logical language with welldefined semantics. For example, suppose we use our system to recommend novels to users based on bibliographic information about the novels and previous usernovel interactions. Imagine that the system recommends the novel "Pride and Prejudice" to user Alex. This prediction can be explained by the Datalog rule “Recommend(x, y) : RomanticGenre(x) & Liked(y,z) & RomanticGenre(z)” (recommend a romantic novel if the user already liked a romantic novel) if the input contains facts such as RomanticGenre(PrideAndPrejudice), RomanticGenre(TheNotebook), and Liked(Alex,TheNotebook). The first aim of this project is to prepare a new suite of benchmarks for our system, by finding and preprocessing relevant data sources. The student will then train and test our GNNbased system on these benchmarks, and evaluate the quality of the computed explanations. If time permits, the student will also investigate how to adjust our system's rule extraction algorithm to produce more useful explanations. Prerequisites: Experience with Python and familiarity with firstorder logic is recommended. Experience with Knowledge Graphs is optional, but desirable. 

Automated Synthesis of Norms in MultiAgent Systems  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  Norms have been widely proposed to coordinate and regulate multiagent systems (MAS) behaviour. In this project we want to consider the problem of synthesising and revising the set of norms in a normative MAS to satisfy a design objective expressed in logic. We focus on dynamic norms, that is, that allow and disallow agent actions depending on the history so far. In other words, the norm places different constraints on the agents' behaviour depending on its own state and the state of the underlying MAS. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Under suitable assumptions, devise and implement norm synthesis techniques against given logical specifications. Reactive Synthesis, GameTheoretic Techniques, Temporal Logics, ATL, ATL*, Strategy Logics, Reasoning about Actions, Planning, Model Checking of MultiAgent Systems Natasha Alechina, Giuseppe De Giacomo, Brian Logan, Giuseppe Perelli: Automatic Synthesis of Dynamic Norms for MultiAgent Systems. KR 2022 (and reference there in)  
Exploring Large Language Models for Reactive Synthesis Problems  Giuseppe De Giacomo, Antonio Di Stasio  Data, Knowledge and Action  C  MSc  Reactive synthesis represents the culmination of declarative programming. By focusing on what a system should accomplish rather than how to achieve it we are able, on the one hand, to simplify the system design process while avoiding human mistakes and, on the other hand, to allow an autonomous agent to selfprogram itself just from highlevel specifications. Linear Temporal Logic (LTL) or LTL on finite traces (LTLf) synthesis is one of the most popular variants of reactive synthesis, being the problem of automatically designing correctbyconstruction reactive systems with the guarantee that all its behaviours comply with desired dynamic properties expressed in LTL/LTLf. Recently, there has been a growing interest in applying Large Language Models (LLMs) to various problems in computer science such as Automated Reasoning, Knowledge Representations and Planning, and Formal Methods in general. Emerging studies have explored the capabilities of LLMs in these fields. The objective of this project is to investigate the feasibility and effectiveness of using LLMs for reactive synthesis problems with specifications expressed in LTL and LTLf. Focus: • Insights into the potential and limitations of LLMs for reactive synthesis with LTL and LTLf specifications. • An LLMguided reactive synthesis framework that integrates natural language processing into the synthesis process. • Empirical evaluation of the performance and effectiveness of LLMbased synthesis algorithms. References: • Amir Pnueli The Temporal Logic of Programs. FOCS 1977: 4657 • Amir Pnueli, Roni Rosner: On the Synthesis of a Reactive Module. POPL 1989: 179190 • Giuseppe De Giacomo, Moshe Y. Vardi: Linear Temporal Logic and Linear Dynamic Logic on Finite Traces. IJCAI 2013: 854860 • Giuseppe De Giacomo , Moshe Y. Vardi: Synthesis for LTL and LDL on Finite Traces. IJCAI • Philipp J. Meyer, Salomon Sickert, Michael Luttenberger: Strix: Explicit Reactive Synthesis Strikes Back! CAV (1) 2018: 578586 • Marco Favorito, Shufang Zhu. LydiaSyft: A Compositional Symbolic Synthesizer for LTLf Specifications • Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, Subbarao Kambhampati: On the Planning Abilities of Large Language Models  A Critical Investigation. CoRR abs/2305.15771 (2023) Prerequisites: Foundations of SelfProgramming Agents and ComputerAided Formal Verification not required but will help 

Nondeterministic Situation Calculus  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  Background and focus The standard situation calculus assumes that atomic actions are deterministic. But many domains involve nondeterministic actions. The key point when dealing with nondeterministic actions is that we need to clearly distinguish between choices that can be made by the agent and choices that are made by the environment, i.e., angelic vs. devilish nondeterminism. “Nondeterministic Situation Calculus” is an extension to the standard situation calculus that accommodates nondeterministic actions and preserves Reiter’s solution to the frame problem. It allows for answering projection queries through regression. But it also provides the means to formalize, e.g., (firstorder) planning in nondeterministic domains and to define execution of ConGolog highlevel program in nondeterministic domains. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Research topic/question and expected contribution Under suitable assumptions (e.g., propositional or finiteobject cases), devise and implement reasoning about action and strategic reasoning techniques for nondeterministic situation calculus based on formal methods methodologies Method Reasoning about Actions, Planning, Model Checking, reactive synthesis Existing work and references Giuseppe De Giacomo, Yves Lespérance: The Nondeterministic Situation Calculus. KR 2021: 216226 Ray Reiter: Knowledge in Action: Logical Foundations for Specifying and Implementing Dynamical Systems, MIT Press 2001 Giuseppe De Giacomo, Yves Lespérance, Hector J. Levesque: ConGolog, a concurrent programming language based on the situation calculus. Artif. Intell. 121(12): 109169 (2000) Giuseppe De Giacomo, Yves Lespérance, Fabio Patrizi: Bounded situation calculus action theories. Artif. Intell. 237: 172203 (2016)  
Planning for Temporally Extended Goals in Linear Time Logics of Finite Traces  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  Planning for temporally extended goals expressed in Linear Time Logics on finite traces requires synthesizing a strategy to fulfil the temporal (process) specification expressed by the goal. To do so one can take advantage of techniques developed in formal methods, based on reduction to twoplayers game and forms model checking. It is crucial to keep in mind that planning is always performed in a domain, which can be thought of a specification of the possible environment reactions. Typically, such a domain specification is is Markovian, i.e., the possible reactions of the environment depend on the current state and the agent action only. However, forms of nonMarkovian domains are also possible and, in fact, of great interest. Note that the domain can be fully observable, partially observable, or nonobservable at all. Each of these settings leads to different forms of planning and different techniques for solving it. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Under suitable assumptions, devise and implement techniques for synthesizing strategies that fulfil temporally extend goals in a planning domain, based on both AI and formal methods methodologies Reasoning about Actions, Planning, Model Checking, reactive synthesis Giuseppe De Giacomo, Sasha Rubin: AutomataTheoretic Foundations of FOND Planning for LTLf and LDLf Goals. IJCAI 2018: 47294735 Giuseppe De Giacomo, Moshe Y. Vardi: LTLf and LDLf Synthesis under Partial Observability. IJCAI 2016: 10441050 Giuseppe De Giacomo, Moshe Y. Vardi: Synthesis for LTL and LDL on Finite Traces. IJCAI 2015: 15581564 Giuseppe De Giacomo, Moshe Y. Vardi: Linear Temporal Logic and Linear Dynamic Logic on Finite Traces. IJCAI 2013: 854860 Giuseppe De Giacomo, Moshe Y. Vardi: AutomataTheoretic Approach to Planning for Temporally Extended Goals. ECP 1999: 226238  
Reactive Program Synthesis Under Environment Specifications in Linear Time Logics on Finite and Infinite Traces.  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  “Reactive synthesis” is the automated synthesis of programs for interactive/reactive ongoing computations (protocols, operating systems, controllers, robots, etc.). This kind of synthesis is deeply related to planning in nondeterministic environments. In this project we are interested in Reactive Synthesis under Environment Specifications, where the agent can take advantage of the knowledge about the environment's behaviour in achieving its tasks. We consider how to obtain the game arena out of the LTL and LTLf specifications of both the agent task and the environment behaviours, and how to solve safety games, reachability games, games for LTLf objectives, and for objectives expressed in fragments of LTL. We also want to understand how winning regions of such arenas are related to the notion of “being able” to achieve desired properties (without necessarily committing to a specific strategy for doing so). We focus on agent tasks that eventually terminate and hence are specified in LTLf. While for the environment we focus on safety specifications and limited forms of guarantee, reactivity, fairness, and stability. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Under suitable assumptions, devise and implement reactive synthesis techniques for fulfil agent task in environment with specified behaviour. Reactive Synthesis, Temporal Logics, 2Player games, Reasoning about Actions, Planning, Model Checking Benjamin Aminof, Giuseppe De Giacomo, Aniello Murano, Sasha Rubin: Planning under LTL Environment Specifications. ICAPS 2019: 3139 Giuseppe De Giacomo, Antonio Di Stasio, Moshe Y. Vardi, Shufang Zhu: TwoStage Technique for LTLf Synthesis Under LTL Assumptions. KR 2020: 304314 Giuseppe De Giacomo, Moshe Y. Vardi: Synthesis for LTL and LDL on Finite Traces. IJCAI 2015: 15581564 Giuseppe De Giacomo, Moshe Y. Vardi: Linear Temporal Logic and Linear Dynamic Logic on Finite Traces. IJCAI 2013: 854860 Benjamin Aminof, Giuseppe De Giacomo, Sasha Rubin: BestEffort Synthesis: Doing Your Best Is Not Harder Than Giving Up. IJCAI 2021: 17661772 Shufang Zhu, Giuseppe De Giacomo: Synthesis of Maximally Permissive Strategies for LTLf Specifications. IJCAI 2022: 27832789 Shufang Zhu, Giuseppe De Giacomo: Act for Your Duties but Maintain Your Rights. KR 2022 Giuseppe De Giacomo, Antonio Di Stasio, Francesco Fuggitti, Sasha Rubin: PurePast Linear Temporal and Dynamic Logic on Finite Traces. IJCAI 2020: 49594965  
Reactive Program Synthesis and Planning under Multiple Environments  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  In this project we consider an agent that operates, with multiple models of the environment, e.g., two of them: one that captures expected behaviours and one that captures additional exceptional behaviours. We want to study the problem of synthesizing agent strategies that enforce a goal against environments operating as expected while also making a best effort against exceptional environment behaviours. We want to formalize these concepts in the context of lineartemporal logic (especially on finite traces) and give algorithms for solving this problem. More generally, we want to formalize and solve synthesis in the case of multiple, even contradicting, assumptions about the environment. One solution concept is based on ``besteffort strategies'' which are agent plans that, for each of the environment specifications individually, achieve the agent goal against a maximal set of environments satisfying that specification. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Under suitable assumptions, devise and implement reactive synthesis techniques that are effective under multiple environments. Reactive Synthesis, Temporal Logics, 2Player games, Reasoning about Actions, Planning, Model Checking Benjamin Aminof, Giuseppe De Giacomo, Alessio Lomuscio, Aniello Murano, Sasha Rubin: Synthesizing Besteffort Strategies under Multiple Environment Specifications. KR 2021: 4251 Benjamin Aminof, Giuseppe De Giacomo, Alessio Lomuscio, Aniello Murano, Sasha Rubin: Synthesizing strategies under expected and exceptional environment behaviors. IJCAI 2020: 16741680 Daniel Alfredo Ciolek, Nicolás D'Ippolito, Alberto Pozanco, Sebastian Sardiña: MultiTier Automated Planning for Adaptive Behavior. ICAPS 2020: 6674 Benjamin Aminof, Giuseppe De Giacomo, Sasha Rubin: BestEffort Synthesis: Doing Your Best Is Not Harder Than Giving Up. IJCAI 2021: 17661772 Benjamin Aminof, Giuseppe De Giacomo, Aniello Murano, Sasha Rubin: Planning under LTL Environment Specifications. ICAPS 2019: 3139 Paolo Felli, Giuseppe De Giacomo, Alessio Lomuscio: Synthesizing Agent Protocols from LTL Specifications Against Multiple PartiallyObservable Environments. KR 2012 Giuseppe De Giacomo, Moshe Y. Vardi: Synthesis for LTL and LDL on Finite Traces. IJCAI 2015: 15581564 Giuseppe De Giacomo, Moshe Y. Vardi: Linear Temporal Logic and Linear Dynamic Logic on Finite Traces. IJCAI 2013: 854860  
Reinforcement learning under safety nonMarkovian Safety Specifications and Rewards expressed in Linear Temporal Logics on Finite Traces  Giuseppe De Giacomo  Data, Knowledge and Action  C  MSc  In some cases, the agent has a simulator of the environment instead of a formal specification, so it needs to learn its strategies to achieve its task in the environment. Sometimes even the task is only implicitly specified through rewards. The key issue is that the type of properties we are often interested in are nonMarkovian (e.g., specified in LTL or LTLf), and hence we need to introduce nonMarkovian characteristics in decision processes and reinforcement learning. A particular promising direction is when such nonMarkovian characteristics can be expressed in Pure Past Linear Temporal Logics. Having taken the Foundation of SelfProgramming and ComputerAided Formal Verification is not required but will be of help. Under suitable assumptions, devise and implement reinforcement learning techniques that remain safe wrt safety specification and achieve rewards specified in linear temporal logics on finite traces Reinforcement Learning, MDP, NonMarkovian Rewards, NonMarkovian Decision Processes, Linear Temporal Logics Ronen I. Brafman, Giuseppe De Giacomo, Fabio Patrizi: LTLf/LDLf NonMarkovian Rewards. AAAI 2018: 17711778Synthesis for LTL and LDL on Finite Traces. IJCAI 2015: 15581564 Giuseppe De Giacomo, Luca Iocchi, Marco Favorito, Fabio Patrizi: Foundations for Restraining Bolts: Reinforcement Learning with LTLf/LDLf Restraining Specifications. ICAPS 2019: 128136 Mohammed Alshiekh, Roderick Bloem, Rüdiger Ehlers, Bettina Könighofer, Scott Niekum, and Ufuk Topcu. Safe reinforcement learning via shielding. AAAI18: 2669–2678. Giuseppe De Giacomo, Marco Favorito, Luca Iocchi, Fabio Patrizi: Imitation Learning over Heterogeneous Agents with Restraining Bolts. ICAPS 2020: 517521 Giuseppe De Giacomo, Marco Favorito, Luca Iocchi, Fabio Patrizi, Alessandro Ronca: Temporal Logic Monitoring Rewards via Transducers. KR 2020: 860870 Giuseppe De Giacomo, Antonio Di Stasio, Francesco Fuggitti, Sasha Rubin: PurePast Linear Temporal and Dynamic Logic on Finite Traces. IJCAI 2020: 49594965 Giuseppe De Giacomo, Moshe Y. Vardi: Linear Temporal Logic and Linear Dynamic Logic on Finite Traces. IJCAI 2013: 854860  
Complexity of popularity in hedonic games  Edith Elkind  Artificial Intelligence and Machine Learning  B  C  MSc  Coalitions and coalition formation are central concerns in the study of multiagent systems as well as cooperative game theory. Typical realworld examples include individuals joining clubs or societies such as music groups or sport teams, task allocation, or students living in shared housing. The formal study of coalition formation is often realized using socalled hedonic games, which originate from economic theory and focus on coalition structures (or partitions) that satisfy various desiderata based on the agents’ preferences over coalitions. One such desideratum is the notion of popularity:A coalition structure S is said to be popular if, for every other coalition structure S', the number of agents preferring S to S' is larger than the number of agents preferring S' to S. In other words, a popular state is a status quo that cannot be replaced by proposing a different solution that wins a majority vote against the status quo. Popular coalition structures seem to be desirable, but, unfortunately, there exist hedonic games, for which popular outcomes do not exist. The concept of popularity gives rise to various interesting computational problems, most importantly the existence problem: given a hedonic game, does it contain a popular coalition structure? Two important classes of hedonic games are additively separable hedonic games and fractional hedonic games. For these classes of games, it is known that the existence problem is NPhard and coNPhard. However, the best known upper bound is that this problem is contained in Sigma_2^p and it is conjectured that the problem is, in fact, Sigma_2^pcomplete. The primary goal of the project is to work on this conjecture. Other possible directions of the project are to consider popularity in restricted classes of hedonic games or to consider related solution concepts, such as strong popularity, mixed popularity, or jointmajority stability. 

Computational analysis of sports ranking systems through game theory  Edith Elkind  Artificial Intelligence and Machine Learning  B  C  MSc  This project proposes a computational game theory analysis focused on addressing the challenge of ranking athletes in sports settings where direct pairwise comparisons may be unavailable, with a particular emphasis on sports like Ironman, Triathlon, and Marathon. Given the inherent difficulty of arranging headtohead competitions among top athletes in such endurance sports, the project seeks to analyze the shortcomings of existing methodologies as well as develop and assess novel ranking systems using insights from computational game theory. The first step in this project will first involve establishing a concrete model and set of assumptions that mirror this realworld problem. Using this model, the project then aims to both theoretically and empirically analyze the current methodology behind the ranking systems being used today. Ultimately, the goal is to provide a provably fair ranking system for such competitions, either based on extensions of existing systems or through a novel ranking concept.  
Strategic nomination in multiwinner voting  Edith Elkind  Artificial Intelligence and Machine Learning  B  C  MSc  In multiwinner voting, each voter specifies their preferences over available candidates (by ranking all candidates, or specifying which candidates they approve), and the goal is to select a fixedsize set of winners. An important phenomenon in reallife voting scenarios is that if voters are not familiar with the candidates, they prefer not to participate in the elections. However, once a voter chooses to participate, they tend to learn more about the candidates and therefore vote according to their true preferences. Thus, a candidate may benefit from nominating another candidate, if they believe that the new candidate's electorate will also support them. The goal of this project is to explore the complexity of the strategic nomination problem, for a variety of voting rules.  
Economic aspects of cybersecurity  Patricia EsteveGonzalez, Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  The links between cybersecurity and economics are multiple and complex, offering topics of discussion at the user, organisational, national, and international levels. The literature covering the overlap of these two fields is limited, and there is a need to cast light on how decisions are taken, how to incentivise the prioritisation of cybersecurity, and how big is the economic incentives behind the protection of digitalised societies. Approaches to the project might involve empirical or theoretical approaches, for example game theory, surveys, interviews, statistical analysis, etc. See, for example: Kianpour, M., Kowalski, S. J., and Øverby, H. 2021. Systematically Understanding Cybersecurity Economics: A Survey. Sustainability, 13(24). https://doi.org/10.3390/su132413677 Hojda, M.H. 2022. Information Security Economics: Cyber Security Threats. Proceedings of the International Conference on Business Excellence, 16(1): 584592. https://doi.org/10.2478/picbe20220056 

Thematic Analysis of National Cybersecurity Maturity Assessments  Patricia EsteveGonzalez, Ioannis Agrafiotis, Louise Axon, Michael Goldsmith, Sadie Creese  Security  B  C  MSc  There is a large investment being made by the international community aimed at helping nations and regions to develop their capacity in cybersecurity. The work of the Global Cyber Security Capacity Centre (based at the Oxford Martin School) studies and documents this: https://www.sbs.ox.ac.uk/cybersecuritycapacity/content/front. There is scope to study in more detail the global trends in capacity building in cybersecurity, the nature of the work and the partnerships that exist to support it. An interesting analysis might be to identify what is missing (through comparison with the Cybersecurity Capacity Maturity Model, a key output of the Centre), and also to consider how strategic, or not, such activities appear to be. An extension of this project, or indeed a second parallel project, might seek to perform a comparison of the existing efforts with the economic and technology metrics that exist for countries around the world, exploring if the data shows any relationships exist between those metrics and the capacity building activities underway. This analysis would involve regression techniques.  
"Beyond MCMC  scalable and approximate Bayesian inference for computational statistics in global health"  Seth Flaxman  Artificial Intelligence and Machine Learning  B  C  MSc  Cosupervisor: Dr Swapnil Mishra (https://smishra.github.io/) In applied work, especially disease modeling, we have reached the limits of what standard MCMC samplers can solve in a reasonable amount of computational time. We will investigate a variety of recently proposed inference schemes, from variational Bayes to deep learning ensemble models to parallel implementations of Sequential Monte Carlo, applying them to popular models in biostatistics, especially multilevel / hierarchical models. The goal will not be just to figure out "what works" but to understand the shortcomings of existing tools, and come up with guidance for practitioners. Cosupervisor 

Disease Mapping with Neural Networks  Seth Flaxman  Artificial Intelligence and Machine Learning  C  MSc  Traditionally, disease mapping has heavily leaned on statistical models such as multivariate normal distributions and Gaussian Processes. However, the everevolving landscape of machine learning, especially in the realm of neural networks, presents an untapped potential for disease mapping. This project aims to harness the latest advancements in spatial data modelling and bring them to the forefront of disease mapping. The goal of the project is to explore cuttingedge techniques, including Bayesian inference with Markov Chain Monte Carlo (MCMC), to unlock new insights and capabilities in disease mapping, bridging the gap between traditional statistical methods and the power of deep learning.  
Active Learning x LLMs  Yarin Gal, Jannik Kossen, Muhammed Razzak  Artificial Intelligence and Machine Learning  MSc  We have a range of projects at the intersection of active learning/subset selection and large language models (LLMs). Active learning is a machine learning approach where the algorithm selectively queries a user or another information source to label new data points with the aim of reducing the amount of data that needs to be labelled for training a model. The motivation behind this approach lies in its efficiency and costeffectiveness, especially in scenarios where labelling data or training is expensive or timeconsuming. It is particularly relevant in the context of LLMs, where finetuning is computationally expensive. The intersection of active learning with LLMs opens new avenues for more efficient model training and finetuning, enabling these models to learn more effectively from smaller, wellchosen subsets of data. This area includes exploring how best to select these subsets, determining which examples are most beneficial for tasks like fewshot learning, and actively selecting demonstrations or test examples to optimize the model's performance and evaluation. This research aims to improve the efficiency and effectiveness of large language models through advanced active learning and subset selection techniques. The key research question investigates how these methodologies can be best applied to reduce training resources while enhancing model performance in diverse NLP tasks. The projects target submission at the major ML conferences. Goals: To be determined based on the project chosen. These projects have been scoped and we will develop a full project plan with the student(s). An expected goal would to be produce research that is of publishable quality at a major ML conference. References:
Prerequisites: Machine Learning and Deep Learning 

Group invariant global pooling  Yarin Gal, Yonatan Gideoni  Artificial Intelligence and Machine Learning  MSc  Background: In GDL often much effort is invested in defining expressive equivariant layers while only simple invariant layers are used. For example, when designing networks over molecular data, there exists a plethora of layers preserving rotations and translations while invariance is attained through pooling. This can restrict the model’s expressivity, making some functions far harder to learn than others. By aggregating over the group’s orbits one can create a rich invariant layer, improving on existing models. Focus: Research questions – how can one design rich invariant layers and (optionally) which functions can they approximate. Expected contribution – building upon and testing an existing invariant architecture. Mostly empirical work, with possibility for theory, depending on the student’s capabilities. Method: The student would chiefly build upon https://arxiv.org/abs/2305.19207. Goals:
Prerequisites: a geometric deep learning course, ideally also some familiarity with group theory 

Learning compatible invariant representations in mixture of experts  Yarin Gal, Yonatan Gideoni  Artificial Intelligence and Machine Learning  MSc  Background: Factored deep mixtures of experts (FDM) can yield a potentially exponential number of network configurations. However, symmetries in the networks’ architectures – such as permuting the weights – can yield features incompatible with later layers. By incorporating symmetry breaking or compatibility blocks that rotate features between layers one can overcome this method’s limitations. Focus: Research questions – how can one get factored deep mixtures of experts to work, specifically regarding training. Expected contribution – developing methods for better training FDMs by accommodating symmetries in the representations. Mostly empirical, consists of implementing and testing simple architectures, with a potential for formalising the approach depending on how independent the student is. Method: The student would build upon https://arxiv.org/pdf/1312.4314.pdf, with input from recent works exploring symmetries in neural architectures such as https://arxiv.org/pdf/2301.12780.pdf, https://arxiv.org/pdf/2106.07682.pdf, and https://arxiv.org/pdf/2209.04836.pdf. Goals:
Prerequisites: a deep learning course with an additional applied one (eg CV/RL/NLP/GDL/etc). 

Temporal feature propagation  Yarin Gal, Yonatan Gideoni  Artificial Intelligence and Machine Learning  MSc  Background: Feature propagation is a technique to impute missing features in graphs. However, many real world graphs are temporal, changing over time. Accordingly, a node that has features at one point may lack them in another. This can be dealt with by treating the problem as a relaxation over the graph+time instead of only the graph. Doing so is relevant also to nonclassical graph problems that contain missing data, eg. noisy video streams. Focus: Research questions – how can one effectively impute features in temporally evolving graphs. Expected contribution – developing, implementing, and testing different methods of temporal feature propagation. Method: The student would build upon https://arxiv.org/abs/2111.12128, with inspiration from other works that treat graphs continuously or analyse them using differential equations, eg. https://arxiv.org/pdf/2202.02296.pdf and https://arxiv.org/abs/2106.10934. Goals:
Prerequisites: a Graph Representation Learning/Geometric Deep Learning course, ideally also a differential equations/computational numerics course 

Topics in Randomised Algorithms and Computational Complexity  Andreas Galanis  Algorithms and Complexity Theory  C  MSc  Description: Andreas Galanis is willing to supervise projects in the areas of randomised algorithms and computational complexity. Problems of interest include (i) the analysis of average case instances of hard combinatorial problems (example: can we satisfy a random Boolean formula?), and (ii) the computational complexity of counting problems (example: can we count approximately the number of proper colourings of a graph G?). The projects would suit mathematically oriented students, especially those with an interest in applying probabilistic methods to computer science.  
Dynamic information design  Jiarui Gan  Artificial Intelligence and Machine Learning  MSc  In information design, a moreinformed player (sender) influences a lessinformed decisionmaker (receiver) by signalling information about the state of the world. The problem for the sender is to compute an optimal signalling strategy, which leads to the receiver taking actions that benefit the sender. Dynamic information design, as a new frontier of information design, generalises the oneshot framework to dynamic settings that are modelled based on Markov decision processes. The goal of the project is to study several variants of the dynamic information design problem and it can be approached from both theoretical and empirical perspectives. Theoretically, the focus is on determining the computational complexity of the optimal information design in different dynamic settings and developing algorithms. A background in computational complexity and algorithm design is beneficial. Practically, the objective is to apply existing algorithms to novel applications, such as traffic management or board games, and to develop algorithms that work effectively in realworld scenarios using stateoftheart methods. Knowledge in Markov Decision Processes, stochastic/sequential games, and Reinforcement Learning is preferred. Related work: J. Gan, R. Majumdar, G. Radanovic, A. Singla. Bayesian persuasion in sequential decisionmaking. AAAI '22 E. Kamenica, M. Gentzkow. Bayesian persuasion. American Economic Review, 2011 S. Dughmi. Algorithmic information structure design: a survey. ACM SIGecom Exchanges 15.2 (2017): 224. Wu, J., Zhang, Z., Feng, Z., Wang, Z., Yang, Z., Jordan, M. I., & Xu, H. (2022). Sequential Information Design: Markov Persuasion Process and Its Efficient Reinforcement Learning. arXiv preprint arXiv:2202.10678. 

Cakecutting with low envy  Paul Goldberg  Algorithms and Complexity Theory  C  MSc  Cakecutting refers to the design of protocols to share a divisible good amongst a collection of agents. A standard startingpoint for cakecutting protocols is the classical "I cut, you choose" rule. This rule is said to be "envyfree" since each player can ensure that they value their own share at least as much as they value the other player's share. Wellknown further work has extended this idea to more than 2 players. In the paper at the URL below, we identify various classes of protocols, and show that one can convert protocols from one kind to another so as to maintain the worstcase level of envy that results. https://arxiv.org/abs/2108.03641 The project is to look at the computational complexity of evaluating the worstcase level of envy that can result from using a given protocol, and related questions about protocols belonging to classes of interest. The project is mainly based on mathematical analysis as opposed to computational experiment. But there is scope for some computational experiment, for example in searching for value functions that result in a high envy.  
Learning probabilistic automata  Paul Goldberg  Algorithms and Complexity Theory  C  MSc  Probabilistic automata are a means of producing randomlygenerated strings that are accepted/recognised by a finite automaton. (They are conceptually similar to hidden Markov models.) The general topic of the project is to find algorithms to reconstruct an automaton based on observations of strings generated by that automaton. It's a topic that has led to a steady stream of papers in the research literature, and the project would focus on one of the many variants of the problem. For example, "labelled automata" in which partial information is provided about the current state of the automaton in the form of random labels associated with states. Another general issue is how to efficiently extract extra information present in long strings generated by the unknown automaton. Within this project topic, there is scope for focusing either on experiments, or on mathematical analysis of algorithms. In the latter case, it would be helpful to know the basic concepts of computational learning theory.  
Synthetic Population Generation  Imran Hashmi, Ani Calinescu  Artificial Intelligence and Machine Learning  C  MSc  Abstract: This project proposes a novel approach of generating synthetic populations of a selected region using Census data (and available statistics). Traditional methods primarily focus on statistical modelling, but these methods lack accuracy and ability to fit the marginals/summary tables. The proposed approach integrates combinatorial optimization for the generation of persons [1] and graph neural networks (GNNs) for predicting household compositions. The integration promises to yield more realistic synthetic populations, vital for numerous applications ranging from urban planning to market research. Motivation: The necessity to create synthetic populations, synthetic mobility, and synthetic environments stems from an urgent need to understand and adapt to the complex dynamics of contemporary urban spaces and establish tools and techniques for nextgeneration urban intelligence [6]. Hence they serve as necessary ingredients for cityscale Digital Twin development. Synthetic populations enable an understanding of demographic distributions and behavioural patterns, serving as a pivotal tool for policy formulation and urban planning. The generation of synthetic mobility patterns, grounded on reliable datasets like travel surveys, offers predictive insights into daily movement patterns, facilitating the optimization of transportation infrastructures and services. Meanwhile, synthetic environments provide a computational playground where urban planners and policymakers can experiment and envisage urban transformations in a controlled, riskfree setting, aiding in the creation of more sustainable and inclusive urban spaces. Together, these synthetic creations forge a datadriven pathway to envisioning, analysing, and moulding future urban landscapes, embodying a convergence of technology and urbanism that holds the potential to revolutionize city planning and development. Related Work: Traditional methods for synthetic population generation, while effective in reproducing statistical summaries, often struggle with capturing complex relationships and generating realistic microlevel data. Mahmood et al., 2023 proposes a novel approach that integrates combinatorial optimization for individual person generation and to predict household compositions [1]. Similar projects like UrbanSim [2] and SynPop2 [3] emphasize microsimulation, while Deep learning approaches like GeoDa explore GNNs for household structures [4]. This project builds on these advancements, offering a potentially powerful framework for future synthetic population generation. Goals and Objectives: Interested students will engage with a research group at the Department of Computer Science, University of Oxford: • To develop a method/extend that synthesizes populations preserving the statistical and relational fidelity of the original Census data. • To extend combinatorial optimization method [1] for individual person generation ensuring maximal attribute representation. • To harness the power of GNNs to predict and design realistic household structures and compositions. • To evaluate the efficacy and accuracy of the proposed method against traditional synthetic population generation techniques. Technical Details: 1.Data Source: The primary data source is Census data. This encompasses demographics, socioeconomic metrics, household compositions, and geographic distributions. 2.Generating Persons using Combinatorial Optimization: • Approach: Adopt combinatorial optimization techniques to select and create individual synthetic persons. The challenge lies in the vast number of potential combinations. • Method: Using optimization algorithms, the aim is to determine the best combination of attributes for each synthetic person to ensure statistical consistency with the original Census data. 3. Generating Household Composition using GNNs and Edge Predictions: • Approach: Recognizing households as graphs, where members are nodes and relationships are edges, we use GNNs to predict the structure and composition of these households. • Method: Train a GNN on existing household data. Posttraining, the network can predict edge formations (relationships) between nodes (persons), thus forming a synthetic household. The edge prediction mechanism helps to ensure realistic household structures. Technical Requirements and Prerequisites: 1.Software & Tools:  Programming environments: Python, TensorFlow or PyTorch for GNNs. 2.Hardware:  Highperformance computing cluster for training GNNs and optimization tasks.  Adequate storage for Census data and synthetic populations. 3. Data:  Complete and cleaned Census data.  Training and test splits for GNNs. 4. Skills and Knowledge:  Familiarity with combinatorial optimization techniques.  Proficiency in deep learning, especially GNNs, Knowledge of edge prediction techniques  Understanding of Census data structure and attributes. Potential Outcomes:
With the everincreasing need for realistic data in research and industry, such innovative approaches are pivotal. This project not only promises advancements in synthetic data generation but also fosters interdisciplinary collaborations between optimization, deep learning, and social sciences. The outcome of this project will provide a foundation for developing robust agentbased models and digital twin at scale. References: [1] Imran Mahmood, Nicholas Bishop, Ioannis Zachos, Anisoara Calinescu, and Michael Wooldridge, "A MultiObjective Combinatorial Optimisation Framework For Largescale Hierarchical Population Synproject", 37th annual European Simulation and Modelling Conference Toulouse, France 2023 [2] UrbanSim: https://cloud.urbansim.com/docs/general/documentation/urbansim.html [3] SynPop2: https://github.com/cran/synthpop/blob/master/R/syn.r [4] GeoDa: http://geodacenter.github.io/ [5] UK Census Data: https://www.nomisweb.co.uk/ [6] Alaa, Ahmed, Boris Van Breugel, Evgeny S. Saveliev, and Mihaela van der Schaar. "How faithful is your synthetic data? samplelevel metrics for evaluating and auditing generative models." In International Conference on Machine Learning, pp. 290306. PMLR, 2022. 

Bioinformatics Projects  Jotun Hein  Computational Biology and Health Informatics  B  C  MSc  Jotun Hein has been in Oxford since 2001 and has supervised many students from Computer Science. His main interest is computational biology and combinatorial optimization problems in this area, especially from phylogenetics, biosequence analysis, population genetics, grammars and the origin of life. Some idea for projects can be obtained by browsing these pages: https://heingroupoxford.com/learningresources/topicsincomputationalbiology/ https://heingroupoxford.com/previousprojects/ A student is also welcome to propose his or her own project. 

Distributional prior in Inductive Logic Programming  Celine Hocquette  MSc  Inductive Logic Programming is a form (ILP) of Machine Learning based on Computational Logic. Given examples and some background knowledge, the goal of an ILP learner is to search for a hypothesis that generalises the examples. However, the search space in ILP is a function of the size of the background knowledge. All predicates are treated as equally likely, and current ILP systems cannot make use of distributional assumptions to improve the search. This project focuses on making use of an assumed given prior probability over the background predicates to learn more efficiently. The idea is to order subsets of the background predicates in increasing order of probability of appearance. This project is a mix of theory, implementation, and experimentation. Prerequisites: logic programming, probability theory, or interest to learn about it  
Identifying relevant background knowledge in Inductive Logic Programming  Celine Hocquette  MSc  Inductive Logic Programming (ILP) is a form of Machine Learning based on Computational Logic. Given examples and some background knowledge, the goal of an ILP learner is to search for a hypothesis that generalises the examples. The search space in ILP is a function of the size of the background knowledge. To improve search efficiency, we want to identify relevant areas of the search space as relevant background knowledge predicates. We propose to evaluate and compare several relevance identification methods such as compression of the examples or statistical approaches. This project is a mix of theory, implementation, and experimentation. Prerequisites: logic programming, statistical learning, or interest to learn about it 

A fast numerical solver for multiphase flow  David Kay  Computational Biology and Health Informatics  B  C  MSc  The AllenCahn equation is a differential equation used to model the phase separation of two, or more, alloys. This model may also be used to model cell motility, including chemotaxis and cell division. The numerical approximation, via a finite difference scheme, ultimately leads to a large system of linear equation. In this project, using numerical linear algebra techniques, we will develop a computational solver for the linear systems. We will then investigate the robustness of the proposed solver. 

Efficient solution of one dimensional airflow models  David Kay  Computational Biology and Health Informatics  B  C  MSc  In this project we will investigate the use of numerical and computational methods to efficiently solve the linear system of equations arising from a one dimensional airflow model within a network of tree like branches. This work will build upon the methods presented within the first year Linear Algebra and Continuous Maths courses. All software to be developed can be in the student’s preferred programming language. General area: Design and computational implementation of fast and reliable numerical models arising from biological phenomena.  
"Extensions of the square root law of steganography"  Andrew Ker  Security  MSc  This is a theoretical project that aims to prove probabilistic results relating to the capacity of different types of cover to conceal a hidden message. It will use tools of probability, information theory, and mathematics. An advanced square root law can be found at http://www.cs.ox.ac.uk/andrew.ker/docs/ADK71B.pdf, and this project either aim for elementary proofs of special cases of that result, or elementary extensions, or to find counterexamples to illustrate that the hypotheses are necessary. Suitable for a student with a background in mathematics who has taken the Probability & Computing option.  
General topics in the area of Steganography and Cryptography  Andrew Ker  Security  B  C  Supervision of general topics in the area of Steganography and Information theory? 

Bots for a Board Game  Stefan Kiefer  Automated Verification  B  C  MSc  The goal of this project is to develop bots for the board game Hex. In a previous project, an interface was created to allow future students to pit their gameplaying engines against each other. In this project the goal is to program a strong Hex engine. The student may choose the algorithms that underly the engine, such as alphabeta search, MonteCarlo tree search, or neural networks. The available interface allows a comparison between different engines. It is hoped that these comparisons will show that the students' engines become stronger over the years. The project involves reading about gameplaying algorithms, selecting promising algorithms and datastructures, and design and development of software (in Java or Scala).  
Small matrices for irrational nonnegative matrix factorisation  Stefan Kiefer  Automated Verification  B  C  MSc  Recently it was shown (https://arxiv.org/abs/1605.06848) that there is a 6x11matrix M with rational nonnegative entries so that for any factorisation M = W H (where W is a 6x5matrix with nonnegative entires, and H is a 5x11matrix with nonnegative entries) some entries of W and H need to be irrational. The goal of this project is to explore if the number of columns of M can be chosen below 11, perhaps by dropping some columns of M. This will require some geometric reasoning. The project will likely involve the use of tools (such as SMT solvers) that check systems of nonlinear inequalities for solvability. Prerequisites: The project is mathematical in nature, and a good recollection of linear algebra is needed. Openness towards programming and the use of tools is helpful.  
Using Virtual Reality to predict how we use memory in natural behaviour: collaborative interdisciplinary projects.  Stefan Kiefer  Automated Verification  B  C  MSc  Name and Email Address of Research Project Supervisor: Dr Dejan Draschkow, dejan.draschkow@psy.ox.ac.uk; Project Description: Although human knowledge and memories represent the past, we form and use them to support future behavior (Nobre & Stokes, 2019). Understanding which factors contribute to learning about the world and successfully finding the learned information in mind is of critical importance for developing methods for supporting this behavior in healthy individuals, but also in individuals with a range of neurocognitive and psychiatric conditions, such as stroke, Alzheimer’s, and Parkinson’s disease. Our novel virtual reality task (Draschkow, Kallmayer, & Nobre, 2021) combines the ecological validity, experimental control, and sensitive measures required to investigate the naturalistic interplay between memory and perception and opens the doors to investigating and supporting complex cognitive functions (https://www.youtube.com/watch?v=GT2kLkCJQbY). In the proposed interdisciplinary projects, computer science and experimental psychology students will be paired to develop and validate sophisticated virtual reality protocols for measuring and supporting complex cognitive mechanism. Specific projects will focus on selected subtopics and vary in scope, depending on students' interests and what kind of project it is (3rd, 4th, or MSc). These include: • Programming and refining gamelike cognitive VR tasks in C# (Unity) • Developing protocols for onlinebased assessments of cognitive functions in C#/JavaScript • Developing algorithms for detecting markers of neurocognitive symptoms (such as tremor for Parkinson’s disease) in multivariate VR data (R/Python) • Developing proofofconcept multimodal (voice, visual, and touch) protocols for supporting learning and memory in VR (with implications for supporting dementia patients) (C#/JavaScript/R/Python) The projects are suitable for students who feel comfortable with highly interdisciplinary work/teams and have experience with (or be open to learn) scientific programming in C#/JavaScript/R/Python. Students will be fully integrated in a successful and collaborative research group and get handson experience with an interactive productdevelopment cycle, including multiple stakeholders. Further related readings are: (Ballard et al., 1997; Hayhoe, 2017; Hayhoe & Ballard, 2014) Relevant readings from cognitive science: Ballard, D. H., Hayhoe, M. M., Pook, P. K., & Rao, R. P. N. (1997). Deictic codes for the embodiment of cognition. In Behavioral and Brain Sciences (Vol. 20, Issue 4, pp. 723–767). https://doi.org/10.1017/S0140525X97001611 Draschkow, D., Kallmayer, M., & Nobre, A. C. (2021). When Natural Behavior Engages Working Memory. Current Biology, 31(4), 869874.e5. https://doi.org/10.1016/j.cub.2020.11.013 Hayhoe, M. M. (2017). Vision and Action. Annual Review of Vision Science, 3(1), 389–413. https://doi.org/10.1146/annurevvision102016061437 Hayhoe, M. M., & Ballard, D. (2014). Modeling task control of eye movements. Current Biology : CB, 24(13), R6228. https://doi.org/10.1016/j.cub.2014.05.020 Nobre, A. C. (Kia), & Stokes, M. G. (2019). Premembering Experience: A Hierarchy of TimeScales for Proactive Attention. Neuron, 104(1), 132–146. https://doi.org/10.1016/j.neuron.2019.08.030  
Quantum Software Projects  Aleks Kissinger  Quantum  C  MSc  I am interested in supervising projects in these areas: (i) developing better quantum compilers for translating highlevel algorithms to real hardware, (ii) simulating complex quantum computations with classical (super) computers, and (iii) automatically verifying the exact or approximate correctness of quantum computations. Many of these projects rely on a mathematical methods based on the ZX calculus or related graphical calculi, mostly developed here in Oxford. Rather than specific, premade projects, I prefer to meet with students and design a project together, often with input/cosupervision from other members of the Quantum Group. Students interested in doing a project are highly encouraged to take the MSc/PartC courses Quantum Processes and Computation and Quantum Software. We also organise various events and group seminars for students doing (or interested in doing) dissertations with the quantum group, usually starting in HT. If you are interested, contact me (aleks.kissinger@cs.ox.ac.uk) and we'll make sure you get informed about those. Research topic  Quantum Software: compiling, classical simulation, and verification 

Computation Theory with Atoms  Bartek Klin  Programming Languages  B  C  MSc  Sets with atoms, also known as nominal sets, are an abstract foundational approach to computing over data structures that are infinite but highly symmetrical, so much so they are finitely presentable and amenable to algorithmic manupulation. Many basic results of classical mathematics and computation theory become more subtle, or even fail, in sets with atoms. For example, in sets with atoms not every vector space has a basis, and a Turing machine may not determinize. A variety of specific topics are available, aimed at mathematically oriented students. Prerequisities: Strong mathematical background is essential. Students who enjoy courses such as "Models of Computation", "Categories, Proofs and Processes" or "ComputerAided Formal Verification" will find this subject suitable. Some relevant literature:


Asymptotically automatic sequences  Jakub Konieczny, James Worrell  Automated Verification  B  C  Asymptotically automatic sequences  that is, sequences whose nth term can be computed by a finite automaton given as input the expansion of n in a given base k  have long been studied in theoretical computer science, number theory, combinatorics and algebra, to name just a few. Recently I introduced a notion tentatively dubbed "asymptotically automatic sequence". This class of sequences remains largely unexplored. It's likely that new results can be obtained by minor modifications of existing arguments, while other cases will present more interesting challenges. The purpose of this project is to see what results on automatic sequences admit a straightforward generalisation to the asymptotic regime, and for which it's possible to construct a counterexample. The goal is to find at least one, preferably more, existing results on automatic sequences and either generalise it to asymptotically automatic sequences, or to disprove such a generalisation. Basic results and definitions can be found in: https://arxiv.org/abs/2305.09885 And a variant of Cobham's theorem can be found in: http://arxiv.org/abs/2209.09588 The standard reference for automatic sequences is “Automatic Sequences: Theory, Applications, Generalizations” by J.P. Allouche and J. Shallit. Prerequisites: Familiarity with automatic sequences / regular languages; basic analysis 

Extensions of Presburger arithmetic by polynomiallike functions  Jakub Konieczny, James Worrell  Automated Verification  B  C  Presburger arithmetic – that is, the first order theory of the natural numbers with addition – is decidable, meaning that there exists a procedure to determine if a given sentence in the language of this theory is true or false. In contrast, Peano arithmetic – which also includes multiplication – is undecidable. This naturally leads to the muchstudied question: Which extensions of Presburger arithmetic are decidable? For instance, adding the square function allows us to define multiplication and hence leads to an undecidable theory. The same applies to other polynomials. But what about more general functions? We recently noticed that for a number of simple generalised polynomials – that is, expressions built up from the usual polynomials using addition, multiplication and the integer part function – the corresponding extensions are also undecidable. The idea behind this project is to investigate more broadly the question of decidability of extensions of the Presburger arithmetic by generalised polynomials. The purpose of this project is to apply existing techniques to show in concrete cases that the extension of the Presburger arithmetic by a given generalised polynomial is undecidable. More generally, one can consider other classes of sequences which have polynomiallike behaviour in a suitable sense. The goal is to show undecidability of the extension of the Presburger arithmetic by at least one generalised polynomial which was not previously covered. More ambitious variants correspond to dealing with wider classes of generalised polynomials. Background on generalised polynomials can be found in (the introductory sections of) the papers: A canonical form and the distribution of values of generalized polynomials by A. Leibman Distribution of values of bounded generalized polynomials by V. Bergelson and A. Leibman Prerequisites: A course in logic, some analysis preferred but not required 

Truthful scheduling for graphs  Elias Koutsoupias  Algorithms and Complexity Theory  B  C  MSc  The aim of the project is to advance our understanding of the limitations of mechanism design for the scheduling problem, the "holy grail" of algorithmic mechanism design. Specifically, we will consider the graphical scheduling problem, in which every task can be only allocated to two machines, and study the approximation ratio of mechanisms for this setting. The aim is to prove both lower and upper bounds. Both directions are hard problems, and we plan to try to gain some understanding by experimentally searching for lower bounds or trying to verify potentially useful statements about the upper bound. Of particular importance is the case of trees and their special case of stars, i.e., when every task can be given either to the root or to a particular leaf. We plan to study not only the standard objective of the makespan, but the more general class of objectives in which the mechanism minimizes the L^p norm of the times of the machines. The case L^infinity is to minimize the makespan, L^1 is to minimize the welfare, and the case L^0 corresponds to the Nash Social Welfare problem, all of which are interesting problems. Further possible directions include considering fractional scheduling and mechanisms without money. Bibliography: George Christodoulou, Elias Koutsoupias, Annamária Kovács: Truthful Allocation in Graphs and Hypergraphs. ICALP 2021: 56:156:20 (https://arxiv.org/abs/2106.03724)  
Truthful scheduling for graphs  Elias Koutsoupias  Algorithms and Complexity Theory  B  C  MSc  The aim of the project is to advance our understanding of the limitations of mechanism design for the scheduling problem, the "holy grail" of algorithmic mechanism design. Specifically, we will consider the graphical scheduling problem, in which every task can be only allocated to two machines, and study the approximation ratio of mechanisms for this setting. The aim is to prove both lower and upper bounds. Both directions are hard problems, and we plan to try to gain some understanding by experimentally searching for lower bounds or trying to verify potentially useful statements about the upper bound. Of particular importance is the case of trees and their special case of stars, i.e., when every task can be given either to the root or to a particular leaf. We plan to study not only the standard objective of the makespan, but the more general class of objectives in which the mechanism minimizes the L^p norm of the times of the machines. The case L^infinity is to minimize the makespan, L^1 is to minimize the welfare, and the case L^0 corresponds to the Nash Social Welfare problem, all of which are interesting problems. Further possible directions include considering fractional scheduling and mechanisms without money. Bibliography: George Christodoulou, Elias Koutsoupias, Annamária Kovács: Truthful Allocation in Graphs and Hypergraphs. ICALP 2021: 56:156:20 (https://arxiv.org/abs/2106.03724)  
Probabilistic Modelling Checking  Marta Kwiatkowska  Automated Verification  B  C  Professor Marta Kwiatkowska is happy to supervise projects involving probabilistic modelling, verification and strategy synthesis. This is of interest to students taking the Probabilistic Model Checking course and/or those familiar with probabilistic programming. Below are some concrete project proposals, but students’ own suggestions will also be considered:


Safety Assurance for Deep Neural Networks  Marta Kwiatkowska  Automated Verification  B  C  Safety Assurance for Deep Neural Networks Professor Marta Kwiatkowska is happy to supervise projects in the area of safety assurance and automated verification for deep learning, including Bayesian neural networks. For recent papers on this topic see http://qav.comlab.ox.ac.uk/bibitem.php?key=WWRHK+19, http://qav.comlab.ox.ac.uk/bibitem.php?key=RHK18 and http://qav.comlab.ox.ac.uk/bibitem.php?key=CKLPPW+19, and also https://www.youtube.com/watch?v=XHdVnGxQBfQ. Below are some concrete project proposals, but students’ own suggestions will also be considered:


An Autopsy Windows registry analysis plugin  Harjinder Lallie, Michael Goldsmith  Security  B  You will develop an Autopsy Forensic Browser plugin. The report will highlight all installed programmes with dates, user account data, other important temporal information. On the face of it, this is a reasonably straightforward task, however we will seek to enrich the functionality for example: 1) a comparative analysis of a windows 10 and the brand new windows 11 registry. 2) identify programmes that were once installed on the machine. 3) provide intelligence on devices that were once attached. Project preparation: If you are considering adopting this project, you should spend some time learning Netbeans IDE, learning how to compile Autopsy Forensic Browser, and strengthening your Java skills. References Singh, A., Venter, H.S. and Ikuesan, A.R., 2020. Windows registry harnesser for incident response and digital forensic analysis. Australian Journal of Forensic Sciences, 52(3), pp.337353. Shaaban, A. and Abdelbaki, N., 2018. Comparison study of digital forensics analysis techniques; Findings versus resources. Procedia Computer Science, 141, pp.545551. 

An Autopsy drone analysis plugin  Harjinder Lallie, Michael Goldsmith  Security  B  Law enforcement investigations increasingly rely on the evidence generated on drone devices. Although there is an abundance of published research on the topic, there are no accepted drone forensics tools. You will develop a drone forensics tool which integrates with Autopsy Forensic Browser. To be able to do that, you will need to use the Netbeans IDE environment and be able to program Java. You will be provided with some code developed by previous project students. The tool you develop {processes, extracts} evidence from {a drone make/model, a small range of 23 make models} which records {Android, IOS, BIN} evidence and {presents all the evidence to the investigator in a graphical format, presents the evidence to the investigator to then be investigated using third part tools}. Elements in curly brackets to be decided. You will be provided with access to a set of drone forensic images. I might (exceptionally) be interested in a nonAutopsy based tool. Project preparation: If you are considering adopting this project, you should spend some time learning Netbeans IDE, learning how to compile Autopsy Forensic Browser, and strengthening your Java skills. References: Bouafif, H., Kamoun, F., Iqbal, F. and Marrington, A., 2018, February. Drone forensics: challenges and new insights. In 2018 9th IFIP International Conference on New Technologies, Mobility and Security (NTMS) (pp. 16). IEEE. AlDhaqm, A., Ikuesan, R.A., Kebande, V.R., Razak, S. and Ghabban, F.M., 2021. Research challenges and opportunities in drone forensics models. Electronics, 10(13), p.1519. 

Applying social network analysis in hard disk based digital investigations: An Autopsy plugin  Harjinder Lallie, Michael Goldsmith  Security  B  Connections between entities in a forensic case can be analysed using social networking theory such as degree centrality, betweenness centrality, closeness centrality, and EigenCentrality. This is a novel way of analysing forensic cases, and can reveal better and more intelligent insights into a case than traditional methods. Although social network analysis has been applied in numerous other fields, there is little research performed within digital investigations of artefacts seized from a crime scene. In this project, you will write a Netbeans/Java based tool which analyses email communications in a forensics case. Using the emails, you will develop social networking graphs at helping forensic investigators identify key persons in a case. Project preparation: If you are considering adopting this project, you should spend some time learning Netbeans IDE, learning how to compile Autopsy Forensic Browser, and strengthening your Java skills. References: Freeman, L., 2004. The development of social network analysis. A Study in the Sociology of Science, 1(687), pp.159167. Scott, J., 2012. What is social network analysis? (p. 114). Bloomsbury Academic. 

Dashcam analysis  Harjinder Lallie  B  C  There exist no current guidelines and very few tools to aid the investigation of a dashcam device, particularly for the purpose of extracting and mapping geospatial data contained therein. This project aims to extract, map and chart geospatial data from a dashcam device and provide insights into the routes and speeds taken by a passenger.You will be provided with a number of dashcam forensic images (.E01 files which can easily be extracted back to MOV/MP4 files). You will be expected to develop the solution using Python, and if possible, integrate the solution with Autopsy, an opensource digital forensic tool. You might consider using an opensource tool such as Exiftool to process the EXIF data. You could choose to take this project to a somewhat different angle by choosing to extract watermark data from the dashcam footage and map that instead. To aid in understanding the background to this project, please see the reference below and the video here: https://warwick.ac.uk/fac/sci/wmg/mediacentre/events/dfls/previousevents/dashcamforensics/ Lallie, H.S., 2020. Dashcam forensics: A preliminary analysis of 7 dashcam devices. Forensic Science International: Digital Investigation, 33, p.200910. Prerequisite. You may want to use tools such as EXIFTOOL, an opensource EXIF data extraction tool. Although EXIFTOOL might serve your needs, you will most probably want to develop an excellent knowledge of the EXIF standard. Additional support can be provided by providing you with access to specific elements of my digital forensics course at the University of Warwick in the form of recorded lectures. That will comprise around 10 hours of learning. I will also provide you with sufficient background in the topic prior to you beginning the study. You will need good programming skills, preferably in Python. 

Keyword searching audio/video files  Harjinder Lallie  B  C  Digital forensic investigators often search for the existence of keywords on hard disks or other storage medium. Keywords are easily searchable in PDF/word/ text/other popular formats, however, current digital forensic tools do not allow for keyword searching through movies/audio. This is essential in cases which involve dashcam footage, recorded conversations etc. The aim of this project is to produce a tool which autotranscribes audio data and then performs a keyword search on the transcribed file – pinpointing the point(s) in the file where the keyword(s) appear. You will be expected to develop the solution using Python, and if possible, integrate the solution with Autopsy, an opensource digital forensic tool. Prerequisite. Additional support can be provided by providing you with access to specific elements of my digital forensics course at the University of Warwick in the form of recorded lectures. That will comprise around 10 hours of learning. You are likely to use the Python SpeechRecognition and possibly the PyAudio libraries. For convenience and demonstration of practicality, you may want to integrate the solution with the opensource forensics tool – Autopsy – and hence will need to develop a good understanding of this tool particularly the keyword search facility.  
Reasoning with Causality in Digital Forensic Cases.  Harjinder Lallie  B  C  Unlike traditional 'wet' forensics, digital forensics is heavily reliant on circumstantial evidence. Quite often a digital investigator is tasked with outlining the sequence of events that took place on a computer system to establish causality and a chain of events. These event sequences are reasonably straightforward to map conceptually, however, there is no solution which 'autolinks' these events. This project applies probabilistic techniques using an attack tree (or alternative) methodology to determining probable event sequences from sequence fragments available on a computer system. You will be supplied with a digital forensic image of a used hard disk, or you may choose to develop your own. You will most likely organize event sequences into categories (e.g. software installation, execution, file download, browsing etc) and then test your proposals on further hard disk images and possibly supplement your findings through qualitative feedback from law enforcement experts (I can connect you). Prerequisite. Knowledge of Autopsy, an opensource forensic tool. Additional support will be given, if required, by providing access to my online wiki selfguided tutorials and by providing access to specific elements of my digital forensics course at the University of Warwick in the form of recorded lectures. That will comprise around 10 hours of learning. I will also provide you with sufficient background in the topic prior to you beginning the study. You will need good programming skills, preferably in Python.  
Augmenting Biometric Datasets with Autoencoders  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab User authentication is of great importance in the age of digital services, but passwords don't work well at scale. Biometrics are increasingly seen as the solution, especially behavioural biometrics that extract user traits implicitly while the user performs other tasks; examples include gait, keystroke dynamics, and responsive eye movements. A major hindrance in behavioural biometric authentication systems is the initial construction of the user template, called the enrolment phase, where the user must laboriously perform the task 'n' times so that their identifying features can be extracted. To facilitate the realworld adoption of these systems, n must be minimised. In recent work, we conducted a user study in which users made mobile payments using both a smartwatch and a smart ring and we collected the inertial sensor (accelerometer, gyroscope, etc.) data from these devices. Using this data, we showed that payment gestures can be used to authenticate the user, both perdevice (watch data can authenticate watch payments and ring data, ring payments) and crossdevice (watch data can authenticate ring payments and vice versa). We also showed that while it was possible to train a reasonable classifier on a small number of samples, those trained on more samples performed proportionally better. The plan for this project would be to use autoencoders to augment a sparse enrolment dataset and show that a model trained on it can perform as well (at differentiating between users) as a model trained on a larger dataset. The project would compare different loss functions and autoencoder variants to find optimum combinations. The dataset contains several possible usecases: the project could focus on either perdevice or crossdevice settings for either device, or a system could be developed on one and tested on another. Prerequisites: Knowledge of data analysis. 

EV Charging Testbed  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab Right now, there is a format war for methods of charging electric vehicles (EVs). One of the most developed systems is the Combined Charging System (CCS). It is underpinned by an international standard called ISO 15118, which describes a complex web of communication that allows vehicle charging to take place. The car talks to the charger it is connected to, in order to report status and place requests for current. The charger, in turn, talks to the backend of the company that runs the charging site (the socalled charging provider), in order to authorise charging sessions and report any errors. It may also talk to a site controller at the same physical location, which is responsible for coordinating multiple chargers and ensuring capacity is shared appropriately between them. The charging provider's backend may federate to other charging providers to permit interoperability between networks. In the near future, it is also expected that the charging provider will be able to talk to energy companies for automated billing, or even to allow electricity to flow the other way  from vehicletogrid (V2G). Exploring the operations and security of this interconnected system is crucially important, while it is still being developed, to avoid building insecure infrastructure that we have to live with for many years to come. A number of problems have already been discovered (see links below) and many more are still to come. Having access to realistic testbeds will accelerate this work substantially, by allowing researchers worldwide to evaluate ideas on a small scale, without having to gain access to costly realworld infrastructure. Our lab has already built a physical testbed for the cartocharger communication, and successfully attacked it. Other researchers have built a system that can simulate multiple such links, but only virtually. Certainly, neither of these existing testbeds implement the full system endtoend, particularly the backend or federation aspects, so it is hard to understand what all the effects of an attack would be. In this project, a student would work towards extending the existing testbed implementations to integrate backend functionality and therefore simulate the whole charging ecosystem. When complete, a user could create a virtual charging provider and register some customers, spawn charging sites with chargers, spawn vehicles  and then simulate full charging sessions. In the ideal case, the implementation would be containerised so that it can be deployed on real hardware, in virtual machines, or a combination of both. It is important to say that there is a lot of scope in this project, and it is not expected that all of it would be implemented in any one year. In fact, there are many parts to suit different interests, across system development and scientific investigation. Meaningful contributions might include: * Build a working, lowpower CCS charger for a toy car * Containerise and integrate several open projects into a connected system * Implement functionality for a charging provider as an opensource reference * Model the PKI operations described in ISO 15118 * Investigate the capability of attackers at various stages of the charging process * Make something crazy, like a cartocar charger with offline crypto payments Some research links: * Our first paper on CCS (testbed in appendix) : https://www.usenix.org/conference/usenixsecurity19/presentation/baker * The virtual testbed from Padua : https://spritz.math.unipd.it/projects/MiniV2G/ * An overview of CCS : https://en.wikipedia.org/wiki/Combined_Charging_System Some open implementations of parts of the ecosystem: * OpenV2G : http://openv2g.org/ * RISEV2G : https://github.com/SwitchEV/RISEV2G * OCPP : https://www.openchargealliance.org/  
Improving Classification of Knocking Gestures  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab In recent work, we conducted a user study in which users wore a smartwatch and a smart ring and knocked on a closed door that had a raspberry pi sensor unit attached to it. We collected inertial sensor (accelerometer, gyroscope, etc.) data from all three devices. We extracted some statistical features from this data and trained random forest classifiers in an attempt to authenticate users by how they knocked on the door. We considered both rhythmic knocking in bursts of 3 and 5 knocks and a secret knock pattern memorised by each user. Given the low entropy in the system, we had expected the classifiers to fail; however, the results showed some promise. It may be possible to achieve better results from this dataset using more sophisticated classification techniques or better feature engineering. Furthermore, we segmented each knock gesture by having the user press a button on the smart ring before and after the gesture to generate bounding timestamps. It may be possible to cut out some noise and improve preprocessing by using peak analysis to segment gestures more precisely, which may also lead to better results. In either case, the plan for this project would be to take the existing dataset and improve on our initial RF approach either by applying better classification, feature engineering, or segmentation techniques. Prerequisites: Knowledge of data analysis. 

Improving Gesture Recognition using Markov Chains  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab In recent work, we conducted a user study in which users made mobile payments using a smartwatch and we collected the inertial sensor (accelerometer, gyroscope, etc.) data from the smartwatch. Using this data, we showed that the tap gesture performed by the user while making such payments can be used to authenticate the user. We also found that the tap gesture is sufficiently distinct from other hand movements that, by recognising it, we can infer the user’s intention to pay (as opposed to it being an accidental payment or skimming attack). We achieved our results by using random forest classifiers on tap gestures represented by sliding windows consisting of between 0.5 and 4 seconds of inertial sensor data. It may be possible to strengthen our intent recognition (gesture recognition) model by treating the tap gesture as a 3tuple of three separate parts, namely: (i) reaching towards the payment terminal, (ii) aligning the watch face with the terminal to establish an NFC connection, and (iii) withdrawing afterwards. The plan would be to attempt to recognise each constituent part separately and use Markov chains to stitch together 3tuples that form a complete gesture. Our dataset will be provided. The research question is whether this approach is more or less effective at recognising tap gestures than our initial RF approach. Prerequisites: Knowledge of data analysis. 

Physically realizable adversarial machine learning models against onscreen classifiers  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab Taking into account the securitycritical contexts in which computer vision models should be used in mixed reality / metaverse, it is of a great interest for us to explore ways of crafting defences against malicious actors using adversarial attacks. Your objective will be to test and improve upon the robustness of existing computer vision models by crafting state of the art adversarial examples in the form of phishing websites that are able to perform on photographs of computer screens captured by a smartphone or XR headset camera.  
Secure Networking in Small Satellite Operating Systems  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab Recent years have seen increased interest in new types of space system, including satellite megaconstellations containing large numbers of satellites, multitenant spacecraft shared between multiple parties, and federated satellite constellations composed of independently operated satellites. Each of these require new security considerations, particularly in the areas of secure communication, key management, and key distribution. In this project a student will extend existing work in key management in new space systems. This could take the form of further developing network simulations to design key management protocols in larger satellite systems. Alternatively, a more implementationbased project is available: extending previous students' work porting the QUIC protocol to the cFS satellite operating system and OPSSAT satellite platform. This would involve implementing and testing further protocols to support more advanced key management functionality. Prerequisites: Students undertaking this project should be familiar with network security concepts and be comfortable writing in C. 

Securing PowerLine Communication  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab PowerLine Communication (PLC) has gained great popularity as a solution for extending network coverage within homes, especially in areas where WiFi signals cannot reach. The primary benefit lies in leveraging the existing electrical wiring infrastructure, eliminating the necessity for extra wiring or specialized network installations. This makes PLC a costeffective and convenient option for deployment. As a result, PLC has also seen a wide adoption in power grid and critical infrastructure applications in recent years. For example, for the interconnection of smart meters. Another example is the Combined Charging System (CCS), one of the most widely adopted DC fast charging standards for Electric Vehicles (EVs), which uses PLC for the communication between the vehicle and the charging station. This communication channel is used to exchange safety critical information, such as battery temperature, maximum charging voltage and current, and state of charge. Unfortunately, our recent research has shown that the PLC communication used by CCS is vulnerable to wireless attacks on the physical layer [1, 2]. We demonstrated that an adversary can eavesdrop on the communication and showed that the charging communication can easily be disrupted. Given the nature of PLC and its tendency to crosstalk, both attacks can be conducted wirelessly and from a distance. In this project, the student will explore PLC security in different contexts, such as smart homes and smart metering infrastructure. Initially, the focus will be on adapting and replicating wireless attacks from previous work. The second phase will involve evaluating new and more advanced attack techniques, along with exploring potential countermeasures. https://github.com/ssloxford/brokenwire https://gitlab.com/rbaker/hpgpemisrx [1] Baker and Martinovic. "Losing the car keys: Wireless PHYlayer insecurity in EV charging." USENIX, 2019. [2] Köhler et al. "Brokenwire: Wireless disruption of ccs electric vehicle charging." Network and Distributed System Security (NDSS) Symposium 2023. Prerequisites: Some familiarity in the area of digital signal processing and with Python. 

Securing Satellite Communication through Transmitter Fingerprinting  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab Due to an increase in the availability of offtheshelf radio hardware, signal spoofing and replay attacks on satellite ground systems have become more accessible than ever. This is particularly a problem for legacy systems, many of which do not offer cryptographic security and cannot be patched to support novel security measures. One method to protect against these attacks is transmitter fingerprinting: identifying transmitters by analysing small differences in the transmitter hardware, expressed in the radio signal. This can be used to distinguish legitimate transmitters from one another, or to identify attackercontrolled transmitters at ground level. In this project a student would build upon existing work in satellite transmitter fingerprinting. This will involve setting up data collection hardware and software to receive and process messages from a satellite constellation, using this data to train machine learning models, and comparing to existing results. See also https://arxiv.org/pdf/2305.06947.pdf (current work) https://github.com/ssloxford/SatIQ Prerequisites: Experience with radio hardware, signal processing, and/or machine learning would be useful for this project. 

Security Monitoring and Control for the cFS Satellite Operating System  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security lab Core Flight System (cFS) is an opensource satellite operating system provided by NASA, providing a standard interface and set of applications which can run on a variety of hardware platforms thanks to its abstraction of the underlying operating systems. Applications and devices within cFS communicate using a publishsubscribe model on a shared bus interface. These messages are not logged or authenticated so a malicious component could affect parts of the satellite it should not have access to. In this project a student would design and implement a security layer over the cFS software bus to provide security monitoring and perapplication access control. Such a system would ideally be backwardscompatible with existing cFS applications. Students undertaking this project should be comfortable writing in C. Useful links:  NASA Core Flight System: https://github.com/nasa/cFS  
Signal Injection Attacks Against Modern Sensors  Ivan Martinovic  Security  B  C  MSc  CoSupervisored by System Security Lab In recent years, the boundaries between the physical and the digital world have become increasingly blurry. Nowadays, many digital systems interact in some way with the physical world. Large and complex cyberphysical systems, such as autonomous and electric vehicles, combine the physical and the digital world and enable the interaction between those two domains. Usually, such systems are equipped with numerous sensors to measure physical quantities, such as temperature, pressure, light, and sound. These physical quantities are vital inputs for the computations and can influence the decisionmaking process of the system. However, the nature of an analog sensor makes it not easily possible to authenticate the physical quantity that triggered a stimulus [1]. For instance, a temperature sensor cannot detect if the stimulus was caused by a legitimate temperature increase or by an adversary using a hairdryer. This is a major concern because the integrity of sensor measurements is critical to ensuring that a system behaves as intended, and a violation of this principle can have serious security, safety and reliability consequences. In our research, we have shown that different sensors are vulnerable to signal injection attacks on the physical layer [2, 3,4]. In this project, a student would analyse the vulnerability of sensors as they are used in modern systems, such as cars, the smart grid and IoT devices. The project will enable the student to research signal injection attacks using different modalities, such as light, acoustic and electromagnetic waves. Moreover, the student will be able to assess the impact of a successful attack against a system as a whole and work on novel countermeasures that can help to improve the security of the next generation of systems. Prerequisites: Some familiarity in the area of digital signal processing and with Python. Useful URLs: https://github.com/ssloxford/ccdsignalinjectionattacks https://github.com/ssloxford/theyseemerollin https://arxiv.org/pdf/2305.06901 References: [1] Kune, Denis Foo, et al. "Ghost talk: Mitigating EMI signal injection attacks against analog sensors." 2013 IEEE Symposium on Security and Privacy. IEEE, 2013.
[3] Köhler, Sebastian, et al. "They See Me Rollin’: Inherent Vulnerability of the Rolling Shutter in CMOS Image Sensors." Annual Computer Security Applications Conference. 2021. [4] Szakály, Marcell, et al. "Assault and Battery: Evaluating the Security of Power Conversion Systems Against Electromagnetic Injection Attacks." arXiv preprint arXiv:2305.06901 (2023).


Transport Layer Security for Satellite Networks  Ivan Martinovic  Security  B  C  MSc  CoSupervisored by System Security Lab Longrange satellite communications networks suffer from high network latencies due to the long distances to satellites. These high latencies have a detrimental effect on the performance of common protocols used for internet traffic, such as the Transmission Control Protocol (TCP). Currently widelyused tools to optimise TCP performance are incompatible with the encrypted traffic of current VPNs. This has led to many operators resorting to providing their communication services unencrypted, leaving customers exposed to eavesdropping attacks. In our research group, we have developed QPEP [1], a novel combination of a VPN and a satellite performanceenhancing proxy, which enables the use of encrypted traffic over satellite links without the usual performance drawbacks. This project would improve on this work in one or more of the following ways: Low Earth Orbit Evaluation Packet Loss Resilience Scalability & Multiuser Environments [1] Pavur, J. C., et al. "QPEP: An actionable approach to secure and performant broadband from geostationary orbit." (2021). https://github.com/ssloxford/qpep https://www.ndsssymposium.org/wpcontent/uploads/2021074paper.pdf  
Using stateoftheart computer vision models to build novel applications on recent XR platforms  Ivan Martinovic  Security  B  C  MSc  Cosupervised by Systems Security Lab In this project, you will be using stateoftheart AI computer vision models that understand their environments (e.g. detect computer screens) and deploying them in novel scenarios on the latest XR (AR/VR) technologies: from NReal glasses, to Microsoft HoloLens, Oculus Quest 2. Your task will be to test and improve upon the robustness of computer vision models by deploying them in novel environments, evaluating their robustness, and building interesting applications around them. For example, this would allow you to build a “physical ad blocker” that would block out any digital advertisement, or privacypreserving remote support application that allows sharing.  
Applied Formal Verification  Tom Melham  Automated Verification  B  C  MSc  I am happy to supervise projects at all levels in applied formal verification. These typically involve handson investigation and innovation to develop new, practical methods of correctness analysis and verification for various kinds of computer systems – either hardware, software, or a combination. They are ideally suited to students who have an interest in computer systems, logic, and handson practical work that has a solid theoretical basis. You don’t have to have taken the course in ComputerAided Formal Verification, if you are willing to learn the theories and technologies required. All my projects are designed to have a strong element of research and sufficient challenge to allow a motivated student to make an excellent contribution. Projects are usually therefore quite ambitious, but they are also designed realistically to fit into the time available. And they always have some fallback options that are less challenging but can still result in an excellent achievement. Rather than offer readymade project ideas, I encourage students with an interest in this area to meet me and together discuss what might align best with their background and interests. I always have several project ideas that link to my current research or research being done by my group. Often projects will have a connection to realworld verification problems in industry, and many of my students will have contact through their projects with leading verification researchers and engineers in industry. If interested to discuss, please contact me. 

Developing computational tools to aid the design of CRISPR/Cas9 gene editing experiments  Peter Minary  Computational Biology and Health Informatics  B  C  MSc  At present, the most versatile and widely used geneediting tool is the CRISPR/Cas9 system, which is composed of a Cas9 nuclease and a short oligonucleotide guide RNA (or guide) that guides the Cas9 nuclease to the targeted DNA sequence (ontarget) through complementary binding. There are a large number of computational tools to design highly specific and efficient CRISPRCas9 guides but there is a great variation in performance and lack of consensus among the tools. We aim to use ensemble learning to combine the benefits of a selected set of guide design tools to reach superior performance compared to any single method in predicting the efficiency of guides (for which experimental data on their efficiency is available) correctly. Recommended for students who has done the Machine Learning and the Probability and Computing courses.


Developing machine learning models for offtarget prediction in CRISPR/Cas9 gene editing  Peter Minary  Computational Biology and Health Informatics  B  C  MSc  The CRISPR/Cas9 gene editing system is composed of a Cas9 nuclease and a short oligonucleotide guide RNA (or guide) that guides the Cas9 nuclease to the targeted DNA sequence (ontarget) through complementary binding but the Cas9 nuclease may also cleave offtarget genomics DNA sequences, which contain mismatches compared to the gRNA, therefore, undesired cleavage could occur. The obvious factors influencing offtarget cleavage activity of the CRISPR/Cas9 gene editing system are the sequence identities of the guide RNA and the offtarget DNA . Various 'basic features' derived from said sequences have been fueling the development of procedural and machine learning models for offtarget cleavage activity prediction but there are numerous 'nonbasic fetures' (such as the sequence context around the offtarget DNA) that may also influence offtarget cleavage activity. The project will aim for the development of novel offtarget clavage activity predictions models using approaches that include but not limited to combining 'basic features' and 'nonbasic features' to increase the accuracy of model predictions of experimental offtarget cleavage activities. Prerequisites: Recommended for students who has done a Machine Learning course and has interest in molecular biology.  
Implementing a Datalog Reasoner  Boris Motik  Data, Knowledge and Action  B  The objective of this project would be for a student to implement a simple Datalog reasoner. Depending on the student’s ambitions, the reasoner could be running in main memory (easier version) or on disk (more challenging). The student would be expected to design the data structures needed to store the data (in RAM or on disk, as agreed with me). On top of this, the student would implement the seminaive algorithm and evaluate the system on a couple of mediumsized datasets. We have in our group readymade datasets that could be used for evaluation. A variation of this project where the student would reuse an existing relational database to store the data. Then, the seminaive algorithm would be implemented either as an extension of the database (assuming the database is open source), or on top of the database (by running SQL queries). This last variant would arguably be much easier as it would involve less design, and more just reusing existing technologies. More advanced students could extend their system to implement various incremental reasoning algorithms. For example, I would give them one of the many papers I’ve written on this topic, and they would have to (a) understand the formal material and (b) implement and evaluate the algorithms. Hence, this project would also give students the opportunity to go as far as they can. Having attended the Databases or Database System Implementation courses, and perhaps to a lesser extent the KRR course, would be a prerequisite for doing this project.  
Implementing a Tableaux Reasoner for Description Logics  Boris Motik  Data, Knowledge and Action  B  The objective of this project would be for a student to implement a simple tableaubased reasoner for description logics. The project would involve the student designing the core data structures, implementing the basic algorithm, realising backtracking, and implementing various optimisations typically used in practical reasoners. The student could use a preexisting API (e.g., the OWL API) to load a description logic ontology, so they could just focus on solving the core problem. The student would be expected to evaluate their implementation on the ontologies from the repository that we’re maintaining in the KRR group (http://www.cs.ox.ac.uk/isg/ontologies/). The project seems to me to provide good opportunity for the student to demonstrate how well they absorbed the CS material (e.g., algorithms, data structures, computational complexity analysis, etc.) taught in our degrees. Also, the project is sufficiently openended so that the student can go quite a long way before running out of options of things to do. Having attended the KRR course would be a prerequisite for doing this project.  
Topics in Automata Theory, Program Verification and Programming Languages (including lambda calculus and categorical semantics)  Andrzej Murawski  Programming Languages  C  MSc  Prof Murawski is willing to supervise in the area of automata theory, program verification and programming languages (broadly construed, including lambda calculus and categorical semantics). For a taste of potential projects, follow this link. 

Concurrent Algorithms and Data Structures  Hanno Nickau  Quantum  C  MSc  Projects in the area of Concurrent Algorithms and Data Structures  
Concurrent Programming  Hanno Nickau  Quantum  B  C  MSc  Projects in the area of Concurrent Programming  
Continual learning by surrogate objectives  Yangchen Pan , Jiarui Gan  Artificial Intelligence and Machine Learning  MSc  A critical challenge in continual learning is catastrophic forgetting (CF), where the acquisition of new information leads to the erosion of previously learned knowledge. This phenomenon poses a substantial barrier, particularly in the context of updating large models, rendering the process computationally unscalable as data increase. CF is primarily attributed to biased gradients resulting from shifts in data sampling distribution over time. This project explores a computationally efficient approach to alleviate forgetting. The core idea involves identifying surrogate objective functions that circumvent the need for extensive memory and computational resources during optimization. The student undertaking this project will work on:


Sample and Computation Efficient Online Adaptation through Offline Reinforcement Learning  Yangchen Pan , Jiarui Gan  Artificial Intelligence and Machine Learning  MSc  In addressing realworld challenges, AI agents, often driven by expansive neural networks like Large Language Models (LLMs) such as GPT, face significant computational and samplerelated costs during training and deployment. Notably, Reinforcement Learning (RL) agents frequently undergo training on vast offline datasets with relaxed computation budgets, followed by deployment or finetuning during an online stage that demands rapid computation. This project aims to investigate methods that capitalize on the disparity between offline and online computation budgets to streamline the training and deployment of RL agents. The student undertaking this project will first delve into relevant literature by studying recommended papers. Subsequently, the student will implement offline RL methods to facilitate downstream online learning. In the third phase, the student will experiment with various online RL algorithm variants. The final algorithm will undergo rigorous empirical testing and comparison to validate its efficiency in handling the challenges posed by large models and varying computation budgets. 

Model checking of POMDPs  David Parker  Automated Verification  B  C  MSc  Formal verification techniques have recently been developed for probabilistic models with partial observability, notably partially observable Markov decision processes (POMDPs), and implemented in software tools such as PRISM. This project will investigate extensions of these techniques, including for example the applicability of sampling based solution methods, adaptation to more expressive temporal logics or extension to partially observable stochastic games.  
Model checking of stochastic games  David Parker  Automated Verification  B  C  MSc  This project will develop formal verification techniques for stochastic games, in particular by considering techniques based on gametheoretic notions of equilibria. A range of such techniques are already implemented in the PRISMgames model checker. This project will consider extensions of these approaches to tackling new types of equilibria, such as Stackelberg equilibria, with potential directions including designing extended temporal logics and solution methods, or modelling new applications, for example from multirobot coordination.  
Probabilistic Model Checking under Uncertainty  David Parker  Automated Verification  B  C  MSc  Formal methods for analysing models such as Markov chains and Markov decision processes can be extended to explicitly reason about model uncertainty, for example by building and analysing interval Markov decision processes. This project will investigate alternative approaches to tackling this problem, which could include alternative models of transition probability uncertainty, factoring in dependencies between different sources of uncertainty, or using bayesian inference to learn model parameters.  
Building databases of mathematical objects in Sagemath (Python)  Dmitrii Pasechnik  Algorithms and Complexity Theory  B  C  MSc 
Prerequisites: Interest in open source software, some knowledge of Python, some maths background. 

Computing with semialgebraic sets in Sage (math)  Dmitrii Pasechnik  Algorithms and Complexity Theory  B  C  MSc  Semialgebraic sets are subsets of R^n specified by polynomial inequalities. The project will extend capabilities of Sage (http://www.sagemath.org) to deal with them, such as CADs computations or sums of squares based (i.e. semidefinite programming based) methods. There might be a possibility for participation in Google Summer of Code (GSoC) or Google Semester of Code with Sage as a GSoC organisation. Prerequisites Interest in open source software, some knowledge of Python, appropriate maths background. 

Automatic translation to GPGPU  Joe PittFrancis  Computational Biology and Health Informatics  B  C  This project involves running cardiac cell models on a highend GPU card. Each model simulates the electrophysiology of a single heart cell and can be subjected to a series of computational experiments (such as being paced at particular heart rates). For more information about the science and to see it in action on CPU see "Cardiac Electrophysiology Web Lab" at https://travis.cs.ox.ac.uk/FunctionalCuration/ An existing compiler (implemented in Python) is able to translate from a domain specific XML language (http://models.cellml.org) into a C++ implementation. The goal of the project is to add functionality to the compiler in order to get OpenCL or CUDA implementations of the same cell models and to thus increase the efficiency of the "Web Lab". 

General GPGPU and high performance computing projects  Joe PittFrancis  Computational Biology and Health Informatics  B  C  I am willing to supervise projects which fit into the general areas of General Purpose Graphics Processing Unit (GPGPU) programming and HighPerformance Computing (HPC). Specific technologies used are likely to based around
All application areas considered although geometric algorithms are favourites of mine. 

General graphics projects  Joe PittFrancis  Computational Biology and Health Informatics  B  C  I am interested in supervising general projects in the area of computer graphics. If you have a particular area of graphicsrelated research that you are keen to explore then we can tailor a bespoke project for you. Specific projects I have supervised in the past include
(Such a project is generally not suitable for MSc students. They should note that in order for this option to work as a potential MSc project then it should be combined with a taughtcourse topic such as machine learning, concurrent programming, linguistics etc.) 

Graphics pipeline animator  Joe PittFrancis  Computational Biology and Health Informatics  B  Prerequisites: Computer graphics, Objectoriented programming The idea behind this project is to build an educational tool which enables the stages of the graphics pipeline to be visualised. One might imagine the pipeline being represented by a sequence of windows; the user is able to manipulate a model in the first window and watch the progress of her modifications in the subsequent windows. Alternatively, the pipeline might be represented by an annotated slider widget; the user inputs a model and then she moves the slider down the pipeline, watching an animation of the process 

Intuitive exploration through novel visualisation  Joe PittFrancis  Computational Biology and Health Informatics  B  C  I am interested in novel visualisation as a way to represent things in a more appealing and intuitive way. For example the Gnome disk usage analyzer (Baobab) uses either a "ring chart" or "treemap chart" Representation to show us which subfolders are using the most disk. In the early 1990s the IRIX file system navigator used a 3D skyscraper representation to show us similar information. There are plenty more ways of representing disk usage: from DAGs to centralised Voronoi diagrams. What kind of representation is most intuitive for finding a file which hogging diskspace and which is most intuitive for helping us to remember where something is located in the filesystem tree? The aim is to explore other places where visualisation gain intuition: for example, to visualise the output of a profiler to find bottlenecks in software, to visual a code coverage tool in order to check that testsuites are are testing the appropriate functionality or even to visualise the prevalence of diabetes and heart disease in various regions of the country.  
The fundamental problem of counterfactual estimation  Francesco Quinzan, Mark van der Wilk  Artificial Intelligence and Machine Learning  MSc  This project deals with the fundamental problem of counterfactual estimation. Counterfactual estimation (CE) refers to the process of estimating or predicting what would have happened in a given situation if different actions or events had taken place. Counterfactual estimation is particularly vital in healthcare due to the complex and interconnected nature of biological systems and the need to understand the true causes behind diseases, treatment effects, and patient outcomes. Although CE holds tremendous promise and potential, estimating counterfactuals remains a significant challenge. Recently, transformertype algorithms have emerged as a powerful tool for the related problem of zeroshot treatment effect estimation. In this project, we will look into transformertype algorithms for CE. The work will consist of: (i) reading and understanding [1]; (ii) extend the framework of [1] to counterfactual estimation from interventional data; (iii) implement and perform experiments with a base model for this extended framework. [1] Towards Causal Foundation Model: on Duality between Causal Inference and Attention. Jiaqi Zhang et al. CoRR abs/2310.00809 (2023) Prerequisites: A student suitable for this project will work with causal inference, attention mechanisms, and possibly the basics of reproducing kernel Hilbert spaces and feature maps. A background on any of these topics is desirable. Coding skills are required.


mmWAVE TECHNOLOGY  Kasper Rasmussen  Security  MSc  One of Stofl’s primary goals is to help facilitate the expansion of the Internet of Things out into the nooks and crannies of the world. The context of this project revolves around our vision for wildlife technology and poaching security. Having secured a partner through the Moholoholo Rehabilitation Centre in South Africa (home to Stoffel the Honey Badger, inspiration for our company, Stofl), we are providing them a LoRaWAN network to deploy these technologies. Together with the Director of the facility and his son, a wellknown antipoacher in the area, discussed a large number of possible LoRa based solutions for the centre, from soil moisture monitoring to feed monitoring, it became clear that the most pressing issue they were facing was regarding poachers breaking in and viciously killing the animals for their skin or tusks. The scope of this project expands on a project that has been started by a student at Imperial College London in an examination into the use of lowpower mmWave sensors to detect people, movement, concealed weapons and detect the difference between them using AI algorithms. The purpose of this project is to design a lightweight, portable and battery operated sensor to assist in antipoaching exercises in South Africa. While the initial solution surrounds the differentiation of various objects, it would need to consider future requirements such as the detection and differentiation between varying weapons made of metal and other materials. Objectives: The study should carry forward the research completed through the previous project with the goal of developing a lightweight, lowpower device capable of detecting the metal casing of a bullet from a minimum range of 2 meters. Once a baseline is set for what is available currently on the market, the project focus should be on optimising the solution by increasing detection range, capabilities and power efficiency. Our overall objective is to provide a framework for potential solutions to be trialed at the Moholoholo Rehabilitation Centre to prevent poachers from killing the animals. Initial objectives involve research and categorisation of what is currently available on the market and comparing them against their inherent limitations. As the project develops, further investigation into potential avoidance methods and security threats of the solutions should be considered. Support: Stofl’s concept is openness and transparency and our schedule is available for anyone to book a meeting with us should a formal time be required. We would invite any interested persons to reach out to us and engage in a discussion and we are available through any communications platforms, whether this be by email, WhatsApp, Zoom, Google Meet, Facetime or face to face. We understand that each individual is unique and we are ready to engage with anyone who we feel is a good fit for us. We will provide regularly scheduled meetings, availability through whatsapp, discord, email, etc. as well as access to our extensive network of partners, companies and experts. Project slides can be viewed here


Statistical shape atlas of cardiac anatomy  Blanca Rodriguez, Pablo Lamata  Computational Biology and Health Informatics  MSc  Description: Cardiac remodelling is the change in shape of the anatomy due to disease processes. 3D computational meshes encode shape variation in cardiac anatomy, and render higher diagnostic than conventional geometrical metrics. Better shape metrics will enable an earlier detection, a more accurate stratification of disease, and more reliable evaluation of remodelling response. This project will contribute to the development of a toolkit for the construction of anatomical atlases of cardiac anatomy, and its translation to clinical adoption. The student will learn about the challenges and opportunities of the crossdisciplinary field between image analysis, computational modelling and cardiology, and be part of a project with a big potential impact on the management of cardiovascular diseases. Prerequisites: Motivation. Good programming skills. Experience with computational graphics and image analysis is an advantage. 

Atmel AVR Assemblers  Alex Rogers  Systems  B  C  MSc  Atmel AVR is a modern RISC instruction set for 8bit microcontrollers. These microcontrollers are relatively simple and make great educational platforms (see for example https://www.cs.ox.ac.uk/people/alex.rogers/2021/03/12/avr_board.html). While the hardware is relatively easy to obtain, there is a lack of development tools that address lowlevel programming in assembly. In this project, you will develop an Atmel AVR assembler targeted at educational users. A range of approaches are possible. One attractive solution would use the WebUSB protocol to communicate directly between a webbased assembler running in the browser and the USB bootloader supported by most Atmel AVR microcontrollers. This would allow direct programming of microcontrollers from the browser, without requiring that users install additional software on their machine. Prerequisites Digital Systems and Computer Architecture useful but not necessary 

EXECUTIVE and WAITLIST for ARM M0+  Alex Rogers  Systems  B  C  MSc  Most modern realtime operating systems for small microcontrollers, such as those using an ARM M0+ processor, allow for a small number of processes which are defined at compile time. In contrast, one of the first examples of a realtime operating system, deployed on the Apollo Guidance Computer (AGC), was more sophisticated. It allowed short WAITLIST tasks to be scheduled at set intervals. Longer EXECUTIVE jobs could be created with different priorities at runtime, typically by the WAITLIST tasks themselves, and these would be completed using whatever spare processor time was available. Famously, during the Apollo 11 landing, the AGC running in the lunar lander ran out of space for new EXECUTIVE jobs, triggering a 1202 alarm. Although low priority jobs updating the astronauts’ displays failed to run, the higher priority jobs controlling the space craft continued as normal and the crew landed safely. This approach to realtime operating system design appears to still offer some advantages. To explore these advantages, in this project, you will develop a realtime operating system for an ARM M0+ processor modelled on this approach. Prerequisites: Digital Systems useful but not essential. 

Resurrecting Extinct Computers  Alex Rogers  Systems  B  C  MSc  While the architecture of current reduced instruction set processors is well established, and relatively static, the early days of computing saw extensive experimentation and exploration of alternative designs. Commercial processors developed during the 1960s, 1970s and 1980s included stack machines, LISP machines and massively parallel machines, such as the Connection Machine (CM1) consisting of 65,536 individual onebit processors connected as a 12dimensional hypercube. This period also saw the development of the first single chip microprocessors, such as the Intel 4004, and the first personal computers, such as the Altair 8800 using the Intel 8080 microprocessor. This project will attempt to resurrect one of these extinct designs (or a scaled down version if necessary) using software simulation or a lowcost fieldprogrammable gate array (FPGA). You will be required research the chosen processor, using both original and modern sources, and then develop a register level description of the device that can be implemented in software or on an FPGA. The final device should be able to run the software of the original. Prerequisites Digital Systems and Computer Architecture useful but not essential. 

Automata Cascades  Alessandro Ronca, Christoph Haase  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  There are several projects available on the theme “Automata Cascades”. Please get in touch for the most uptodate information. Automata Cascades are a modular, structured way to represent automata. An automaton is seen as consisting of many simpler automata that interact with each other. Fundamental results establish that large classes of automata can be expressed as cascades of very simple components. Please see the reference below for an introduction to the formalism. There are projects focusing on: (i) expressivity of classes of automata cascades, (ii) computational complexity of decision problems for automata cascades, (iii) learning automata cascades—complexity results and learning algorithms. The specific topic will be identified together with the candidate, based on the candidate’s interests and background. Reference: Alessandro Ronca, Nadezda Alexandrovna Knorozova, Giuseppe De Giacomo: Automata Cascades: Expressivity and Sample Complexity. AAAI 2023 Prerequisites: Some familiarity with automata. Familiarity with Complexity theory and Computational Learning Theory might be needed depending on the focus of the project 

Implementing a Reasoner for the Transformation Logics  Alessandro Ronca, Christoph Haase  Artificial Intelligence and Machine Learning, Automated Verification  B  The Transformation Logics (TLs) are a recentlyproposed family of logics for capturing patterns over sequences [1]. The logics are expected to have a broad range of applications, but there is no implementation so far. The goal of the project is to implement the first reasoner for the Transformation Logics, and design an experimental evaluation. Relevant paper: [1] The Transformation Logics, Alessandro Ronca, IJCAI 2024 Prerequisites: Some familiarity with Propositional Logic, Temporal Logic, Automata. Some familiarity with Semigroups and Groups would be helpful. 

Neural Networks and Formal Languages  Alessandro Ronca, Christoph Haase  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  There are several projects available on the theme “Neural Networks and Formal Languages”. Please get in touch for the most uptodate information. Neural networks such as RNNs and Transformers are able to process sequences, and hence they can operate as language recognisers. On one hand, this makes them applicable to settings where automata and temporal logics are typically employed. On the other hand, it implies that the capabilities of these networks can be studied through the lens of formal languages and automata theory. The goal of this area of research is to get a better understanding of neural networks operating on sequences, improve their capabilities, and broaden their range of applications. There are projects focusing on: (i) expressivity results, (ii) learning algorithms, (iii) benchmarks. The specific topic will be identified together with the candidate, based on the candidate’s interests and background. This project is at the intersection of two fields, automata and formal languages on one side, and neural networks on the other side. Applicants interested in exploring this intersection should not be discouraged if their background is mostly in only one of the two fields. Reference: Nadezda Alexandrovna Knorozova, Alessandro Ronca: On the Expressivity of Recurrent Neural Cascade, AAAI Prerequisites: Some familiarity with automata and formal languages, or with neural networks. 

NonMarkov Reinforcement Learning  Alessandro Ronca, Christoph Haase  Artificial Intelligence and Machine Learning, Automated Verification  C  MSc  There are several projects available on the theme “NonMarkov Reinforcement Learning”. Please get in touch for the most uptodate information. NonMarkov Reinforcement Learning focuses on how an agent can learn to behave optimally in environments where the history of past observations is relevant to predict what will happen in response to the agent’s actions. It is a more general setting than the one described by Markov Decision Processes (MDPs) where it is sufficient to consider the current observation. The challenge is to develop effective methods to learn and take decisions according to the sequence of past observations. Please see the references below for an introduction to the problem. There are projects focusing on: (i) performance bounds, e.g., PACstyle bounds and regret bounds, (ii) efficient algorithms, (iii) novel approaches to overcome the limitations of existing algorithms, (iv) implementation and experimental evaluation. The specific topic will be identified together with the candidate, based on the candidate’s interests and background. References: Alessandro Ronca, Giuseppe De Giacomo: Efficient PAC Reinforcement Learning in Regular Decision Processes. IJCAI 2021. Roberto Cipollone, Anders Jonsson, Alessandro Ronca, Mohammad Sadegh Talebi: Provably Efficient Offline Reinforcement Learning in Regular Decision Processes. NeurIPS 2023. Prerequisites: Some familiarity with Reinforcement Learning, Automata, and Computational Learning Theory. 

Explainability of Features Learned by Diffusion Models  Christian Rupprecht  Artificial Intelligence and Machine Learning  B  C  MSc  Large scale image diffusion models (such as StableDiffusion) have shown remarkable capabilities in generating and editing images and have transformed the field of generative image modelling. It has been shown that their learned features are very general and can thus be used for many downstream tasks such as segmentation and semantic correspondences. However, much less research has been done in understanding these features. In this project, we will investigate the explainability of these features through the use of typical explainability frameworks and techniques. As these features are of a generative nature, it is possible that they contain quite different information than features from prior discriminative models. Goals: • Extract features from large diffusion models. • Setup explanation methods to utilise these features. • Find similarities and differences between those and features from other models. Stretch Goal: • There are many ways to extract features from diffusion models. Based on the results of the previous goals: is there a principled way to find which is the best one? References: Rombach, Robin, et al. "Highresolution image synthesis with latent diffusion models. 2022 IEEE." CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2021. Zhang, Junyi, et al. "A Tale of Two Features: Stable Diffusion Complements DINO for ZeroShot Semantic Correspondence." arXiv preprint arXiv:2305.15347 (2023). Laina, Iro, Yuki M. Asano, and Andrea Vedaldi. "Measuring the Interpretability of Unsupervised Representations via Quantized Reversed Probing." International Conference on Learning Representations. 2021. Prerequisites: Machine Learning 

Understanding Bias in Object Detection Models  Christian Rupprecht  Artificial Intelligence and Machine Learning  B  C  MSc  Object detection is a fundamental task in computer vision. The goal is to locate and classify individual instances of objects in images (e.g., people, cars, cups, sheep, etc.). Most current models have been trained on benchmark datasets that consist of handannotated images collected from the internet. This introduces bias in the training data which in turn has been shown to lead to biased models. In this project, we will make use of modern image editing techniques, such as inpainting diffusion models, to modify images in specific ways (e.g., changing the apparent age of people, skincolor, daynight, …), which allows us to measure the impact of the edit on the object detection performance. Goals: • Setup an evaluation framework of object detection models that allows modification of the test data. • Use different image editing techniques to edit the test data. • Test several hypotheses for bias in multiple different models. Stretch Goal: • Bias can potentially mitigated by including the modified data during training of the model. References: Rombach, Robin, et al. "Highresolution image synthesis with latent diffusion models. 2022 IEEE." CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2021. Brooks, Tim, Aleksander Holynski, and Alexei A. Efros. "Instructpix2pix: Learning to follow image editing instructions." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. Lin, TsungYi, et al. "Microsoft coco: Common objects in context." Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 612, 2014, Proceedings, Part V 13. Springer International Publishing, 2014. Singh, Krishna Kumar, et al. "Don't judge an object by its context: Learning to overcome contextual bias." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. Prerequisites: Machine Learning 

Theoretical and experimental study of Boolean matrices and automata  Andrew Ryzhikov  B  C  MSc  Let A be a nondeterministic finite automaton (NFA) with n states. Such automaton is called unambiguous if for every word w and every pair p, q of its states, the number of paths from p to q labelled by w is at most one (sometimes such automata are called diamondfree). A word w is called mortal for A if for every state q there is no path starting in q and labelled by w. In https://arxiv.org/abs/1808.00940, Kiefer and Mascle proved that if an unambiguous NFA admits a mortal word, then the length of a shortest such word is at most n^5. However, the best known lower bound for this value is quadratic in n. The experimental part of the project is to perform an exhaustive search for small unambiguous NFAs, and come up with a conjecture of how this upper bound should actually look like. This is challenging, because even for small n the number of NFAs is very large, and finding the length of a shortest mortal word is hard, but there are some good existing approaches to deal with that (for example, https://arxiv.org/abs/2207.05495). The theoretical part is to try to improve the lower or the upper bound for this question, ideally using the examples obtained from the experimental part. As an intermediate step for that, we will look at restricted cases, especially at unambiguous NFAs associated to finite prefix codes (also known as Huffman codes) and special classes of Boolean matrices. Prerequisites: good familiarity with discrete mathematics, algorithms and linear algebra (for both theoretical and experimental parts), ability to design and implement nontrivial algorithms in a programming language of your choice (for the experimental part) 

Categorytheoretic syntactic models of programming languages  Philip Saville, Sam Staton  Programming Languages  C  MSc  Syntactic models  categories constructed from the syntax of a programming language  play a key role in categorytheoretic denotational semantics. Showing such models exist and satisfy a suitable universal property amounts to giving a sound and complete semantic interpretation for the language in question. Often this involves carefully studying the interplay between program features and categorical structures. The three main aims of the project are as follows. Firstly, to construct syntactic models for two idealised effectful functional programming languages, namely Moggi's monadic metalanguage [1] and computational lambda calculus [2]. Next, to prove their universal properties, and finally to use these to give syntactic translations between the languages. The starting point would be to understand the categorical semantics of the simplytyped lambda calculus, the monadic metalanguage and the computational lambda calculus. Extensions would include exploring extensionality / wellpointedness and constructing fully abstract syntactic models of these languages. [1] E. Moggi, "Notions of computation and monads," 1991 [2] E. Moggi, Prerequisite courses:  Categories, proofs and processes Useful background courses:  Principles of programming languages  Lambda calculus and types


Achieving Superalignment through WeaktoStrong Generalization  Christian Schroeder de Witt, Philip Torre, Ani Calinescu  Artificial Intelligence and Machine Learning  B  C  MSc  The capabilities of AI systems have significantly grown within a short amount of time. Further advances in synthetic data generation, selfplay, and other techniques are set to improve their performance further, potentially resulting in superhuman capabilities. This poses an important safety question: But how can humans supervise future systems that are much smarter than themselves? Whereas current systems are aligned using human data, tasks that are too hard to solve for humans will require progress in the nascent field of superalignment [1]. In this project, we will conduct a critical examination of existing superalignment frameworks [2], and leverage latest work in semisupervised learning and other fields in order to advance our understanding of superalignment. This project is designed to lead to publication. We are looking for a highlymotivated student. For this project, we will be able to receive advice from Collin Burns (OpenAI). [1] https://openai.com/blog/introducingsuperalignment [2] https://cdn.openai.com/papers/weaktostronggeneralization.pdf 

Active Defenses against Illusory Attacks  Christian Schroeder de Witt, Philip Torre, Alessandro Abate  Automated Verification  B  C  MSc  Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We recently demonstrated that existing observationspace attacks on reinforcement learning agents have a common weakness: while effective, their lack of informationtheoretic detectability constraints makes them detectable using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. In response, we introduced illusory attacks, a novel form of adversarial attack on sequential decisionmakers that is both effective and of epsilonbounded statistical detectability. In this project, we, for the first time, investigate how victims of epsilonbounded illusory attacks can actively defend themselves, i.e. adapt their testtime behaviour to maximise informationtheoretic detectability. Drawing from the active learning literature, and inspired by Bayesian optimal experiment design and constraint preferences in economics, we will develop novel ways of optimising for the victim’s best response to a costbounded adversary. This project is designed to lead to publication. We are looking for a highlymotivated student with interest in theory. [1] Tim Franzmeyer, Stephen Marcus McAleer, Joao F. Henriques, Jakob Nicolaus Foerster, Philip Torr, Adel Bibi, Christian Schroeder de Witt, Illusory Attacks: Detectability Matters in Adversarial Attacks on Sequential DecisionMakers, https://openreview.net/forum?id=F5dhGCdyYh (under review at ICLR 2024)


Enhancing WorstCase Safety in Large Language Models through Influence Functions and Backdoor Detection  Christian Schroeder de Witt, Philip Torre, Ani Calinescu  Artificial Intelligence and Machine Learning  B  C  MSc  Large language models have advanced in various fields, but concerns about their safety in worstcase scenarios persist [1]. Interpretability methods are used to understand these models' decisions but have limitations — they mainly focus on average cases and lack ground truth [2]. This proposal aims to overcome these limitations by injecting backdoors into language models to provide ground truth and developing an efficient influence functionbased method for backdoor detection. In doing so, we will leverage exciting new progress allowing influence functions to be scaled to large language models [3]. This project is designed to lead to publication. We are looking for a highlymotivated student. We will be able to collaborate with Dr. Ameya Prabhu (Tuebingen University) on this project. [1] Carlini et al., Are aligned neural networks adversarially aligned?. June 2023. http://arxiv.org/abs/2306.15447. arXiv:2306.15447 [cs] [2] Conmy et al., Towards Automated Circuit Discovery for Mechanistic Interpretability, October 2023. http://arxiv.org/abs/2304.14997. arXiv:2304.14997 [cs] [3] Grosse et al., Studying Large Language Model Generalization with Influence Functions, https://arxiv.org/abs/2308.03296, arXiv:2308.03296 [cs.LG] 

Extending AIGenerated Steganography to Generative Diffusion Models  Christian Schroeder de Witt, Philip Torre, Mark van der Wilk  Artificial Intelligence and Machine Learning  B  C  MSc  Steganography is the practice of encoding a plaintext message into another piece of content, called a stegotext, in such a way that an adversary would not realise that hidden communication is occurring. In a recent breakthrough [1,2,3], we showed that messages can be encoded into the output distribution of arbitrary autoregressive neural networks with perfect security [4], i.e. without changing the output distribution, hence rendering the hiding of secret messages informationtheoretically undetectable. In this project, we extend perfectly secure steganography to generative AImodels that are not explicitly representable as discrete autoregressive distributions, such as e.g. diffusion models. Such an extension would allow for applying informationtheoretic steganography to highresolution images, potentially resulting in novel tools for investigative journalism and watermarking. This project is designed to lead to publication. We are looking for a highlymotivated student. [1] https://arxiv.org/abs/2210.14889 [2] https://www.quantamagazine.org/secretmessagescanhideinaigeneratedmedia20230518/ [3] https://www.scientificamerican.com/article/aicouldsmugglesecretmessagesinmemes/ [4] https://www.sciencedirect.com/science/article/pii/S0890540104000409 

Illusory dynamic patch attacks on autonomous driving agents  Christian Schroeder de Witt, Philip Torre, Ivan Martinovic  Security  B  C  MSc  Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We recently demonstrated that existing observationspace attacks on reinforcement learning agents have a common weakness: while effective, their lack of informationtheoretic detectability constraints makes them detectable using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. In response, we introduced illusory attacks, a novel form of adversarial attack on sequential decisionmakers that is both effective and of epsilonbounded statistical detectability. In this project, we will be extending illusory attacks to dynamic patch attacks in autonomous driving settings. To this end, we consider a threat model in which dynamic patches can be projected into the car’s input sensors (be it visual, or using LIDAR spoofing [4]). For example, an adversarial car could project adversarial patches into the car behind it using a digital screen. This could mislead the car behind it into believing that the car in the neighbouring lane is ceding when it is not, thus causing inefficient or dangerous situations. In this project, we will develop novel methods that ensure that the patch attacks generated are illusory in entity space, i.e. the car’s internal dynamics model of the world will perceive these attacks to be temporally and logically consistent  unlike in conventional adversarial attacks. To this end, we will leverage the latest advances in world model learning and reinforcement learning selfplay. Furthermore, we will investigate possible ways of defending against illusory attacks both in entity and input space, building up on our recent work in outofdistribution dynamics detection [2]. We will work with stateoftheart autonomous driving simulators [3], and develop a novel RL benchmark environment suite for adversarial attacks in autonomous driving. If successful, we will then demonstrate our novel dynamic patch attacks in an exemplary realworld setting. This project is designed to lead to publication. We are looking for a highlymotivated student. [1] Tim Franzmeyer, Stephen Marcus McAleer, Joao F. Henriques, Jakob Nicolaus Foerster, Philip Torr, Adel Bibi, Christian Schroeder de Witt, Illusory Attacks: Detectability Matters in Adversarial Attacks on Sequential DecisionMakers, https://openreview.net/forum?id=F5dhGCdyYh (under review at ICLR 2024) [2] Linas Nasvytis, Kai Sandbrink, Jakob Foerster, Tim Franzmeyer and Christian Schroeder de Witt, Rethinking OutofDistribution Detection for Reinforcement Learning: Advancing Methods for Evaluation and Detection, AAMAS 2024 (Oral) [3] https://github.com/waymoresearch/waymax [4] Yulong Cao et al., Adversarial Sensor Attack on LiDARbased Perception in Autonomous Driving, ACM SIGSAC 2019 

Learning MinimumEntropy Couplings using AlphaZero and Geometric Deep Learning  Christian Schroeder de Witt, Philip Torre, Michael Bronstein  Artificial Intelligence and Machine Learning  B  C  MSc  Steganography is the practice of encoding a plaintext message into another piece of content, called a stegotext, in such a way that an adversary would not realise that hidden communication is occurring. In a recent breakthrough [1,2,3], we showed that messages can be encoded into the output distribution of arbitrary autoregressive neural networks with perfect security [4], i.e. without changing the output distribution, hence rendering the hiding of secret messages informationtheoretically undetectable. In this project, we seek to improve the heart of our perfectlysecure encoding algorithm, namely minimum entropy couplings. As constructing exact minimum entropy couplings is known to be NPhard, we currently use an approximation algorithm that runs in loglinear time. For large block sizes, which are crucial to achieving good transmission rates, this algorithm is slow as it is not readily parallelisable. In this project, we will build upon recent advances in deep reinforcement learning selfplay [5] and leverage graph networks to implement constraints on the sparse neural network outputs [6] in order to learn deep minimum entropy coupling networks. If successful, this project will greatly increase the feasibility of AIgenerated steganography on mobile devices, as well as lead to novel methodological approaches for neural network architectures for sparse doublystochastic matrix outputs. In this project, we will be working with stateoftheart MARL implementations in JAX [2]. This project is designed to lead to publication. We are looking for a highlymotivated student. [1] https://arxiv.org/abs/2210.14889 [2] https://www.quantamagazine.org/secretmessagescanhideinaigeneratedmedia20230518/ [3] https://www.scientificamerican.com/article/aicouldsmugglesecretmessagesinmemes/ [4] https://www.sciencedirect.com/science/article/pii/S0890540104000409 [5] https://deepmind.google/discover/blog/alphadevdiscoversfastersortingalgorithms/ [6] Modeling relational data with graph convolutional networks, M Schlichtkrull, TN Kipf, P Bloem, R Van Den Berg, I Titov, M Welling, The Semantic Web: 15th International Conference, ESWC 2018 [7] Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges, Michael M. Bronstein, Joan Bruna, Taco Cohen, Petar Veličković, arXiv:2104.13478 [cs.LG]


MultiAgent Illusory Attacks and Defenses  Christian Schroeder de Witt, Philip Torre, Ivan Martinovic  Security  B  C  MSc  Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We recently demonstrated that existing observationspace attacks on reinforcement learning agents have a common weakness: while effective, their lack of informationtheoretic detectability constraints makes them detectable using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. In response, we introduced illusory attacks, a novel form of adversarial attack on sequential decisionmakers that is both effective and of epsilonbounded statistical detectability. In this project, we extend illusory attacks to decentralised multiagent settings. In particular, we are asking the question how teams of agents can coordinate on implementing illusory attacks on other agents together, and how these can be jointly detected (e.g. in Poker or Autonomous Driving). In this project, we will be working with stateoftheart MARL implementations in JAX [2]. This project is designed to lead to publication. We are looking for a highlymotivated student with interest in theory. [1] Tim Franzmeyer, Stephen Marcus McAleer, Joao F. Henriques, Jakob Nicolaus Foerster, Philip Torr, Adel Bibi, Christian Schroeder de Witt, Illusory Attacks: Detectability Matters in Adversarial Attacks on Sequential DecisionMakers, https://openreview.net/forum?id=F5dhGCdyYh (under review at ICLR 2024) [2] Alexander Rutherford, Benjamin Ellis, Matteo Gallici, Jonathan Cook, Andrei Lupu, Gardar Ingvarsson, Timon Willi, Akbir Khan, Christian Schroeder de Witt, Alexandra Souly, Saptarashmi Bandyopadhyay, Mikayel Samvelyan, Minqi Jiang, Robert Tjarko Lange, Shimon Whiteson, Bruno Lacerda, Nick Hawes, Tim Rocktaschel, Chris Lu, Jakob Nicolaus Foerster, JaxMARL: MultiAgent RL Environments in JAX, https://arxiv.org/abs/2311.10090, accepted at AAMAS 2024 

Preventing Malicious Collusion between Advanced AI Systems  Christian Schroeder de Witt, Philip Torre, Alessandro Abate  Automated Verification  B  C  Consider settings in which one or more principals assign a task to a team of generative AI agents, for example a scheduling or negotiation task. The principals monitor the (not necessarily humanintelligible) communications between the agents and intervene if deemed necessary, hoping to prevent agents from pursuing undesirable joint strategies. In this project, we investigate the question if, when, and how optimisation pressure may lead generative AI agents to hide communications from their principals. We survey the landscape of steganography (information hiding), and, for a given level of security, identify the required knowledge and capabilities of generative AI agents. From these, we design a roadmap for model evaluation building on our recent work in this space [1][2]. We then empirically test the ability of stateoftheart LLMs to engage in different types of covert communication given a variety of optimisation pressures. This project is designed to lead to publication. We are looking for a highlymotivated student. [1] Secret Collusion Among Generative AI Agents: A Model Evaluation Framework, Sumeet Ramesh Motwani, Mikhail Baranchuk, Philip H.S. Torr, Lewis Hammond, and Christian Schroeder de Witt, to appear  also see this talk here: https://www.alignmentworkshop.com/nolatalks/christianschroederdewittperfectlysecuresteganographyandllmcollus [2] https://www.quantamagazine.org/secretmessagescanhideinaigeneratedmedia20230518/ 

Supercharging OutofDistribution Dynamics Detection  Christian Schroeder de Witt, Philip Torre, Varun Kanade  Artificial Intelligence and Machine Learning  B  C  MSc  The challenge in OutofDistribution Dynamics Detection (O2D3) is to test whether a given sequential control environment at testtime is the same as the one the agent was trained in. This capability is central to both AI safety, and security: A robot controller may have systematic errors at testtime, or an adversary may attack the AI agent’s sensors. In both cases, if anomalies in the testtime environment are not detected as soon as possible, potentially catastrophic outcomes may follow. In this project, we are building upon the latest advances in statistics, including doublyrobust estimation techniques, informationtheoretic hypothesis testing and latest advances in outofdistribution detection, in order to surpass the current stateoftheart in O2D3 [1]. This project is designed to lead to publication. We are looking for a highlymotivated student with interest in theory. [1] Linas Nasvytis, Kai Sandbrink, Jakob Foerster, Tim Franzmeyer and Christian Schroeder de Witt, Rethinking OutofDistribution Detection for Reinforcement Learning: Advancing Methods for Evaluation and Detection, AAMAS 2024 (Oral) 

Adapting Red to the Language Server Protocol  Bernard Sufrin  Programming Languages  B  C  The Language Server Protocol is an open, JSONRPCbased protocol for use between source code editors or integrated development environments (IDEs) and servers that provide ”language intelligence tools”: programming languagespecific features like code completion, syntax highlighting and marking of warnings and errors, as well as refactoring routines. The goal of the protocol is to allow programming language support to be implemented and distributed independently of any given editor or IDE. In the early 2020s LSP quickly became a norm for language intelligence tools providers.


Modeless Structure Editing  Bernard Sufrin  Programming Languages  B  C  Oege de Moor and I wrote a paper called Modeless Structure Editing for Tony Hoare’s retirement symposium. What we had in mind was a family of modeless editors whose underlying model of the document being edited is akin to its abstract syntax tree. It’s clear what this means if the document represents, for example, a program in a programming language or a proof in a proof calculus; but we conjecture that structured prose documents will also be amenable to this approach to a useful extent. Editing an abstract syntax tree model does not mean that one must necessarily work with a “treeshaped” view of the document. The benefits of editing an abstract syntax tree model include:
Such an approach has the potential to yield dividends in performance and simplicity of implementation of incremental semantic checks, codegeneration, etc. The first phase of this project will construct (in Scala) a prototype structure editing toolkit implementing the model explained in this paper. Subsequent phases will provide one or two case studies in its use. The architecture of an editor that could use such a document model already exists – in the shape of our AppleRed editor, whose document model is (in principle) “pluggable”, and that currently uses a linear model (document as a sequence of lines of text). What will be needed is to construct a new “plugin” conforming to the AppleRed document model interface, and to adapt and generalise that interface where necessary. 

"Parallel reasoning in Sequoia"  David Tena Cucala, Bernardo Cuenca Grau  Data, Knowledge and Action  C  MSc  Sequoia is a stateoftheart system developed by researchers in our department for automated reasoning in OWL 2, a standard ontology language in the Semantic Web. Sequoia uses consequencebased calculus, a novel and promising approach towards fast and scalable reasoning. Consequencebased algorithms are naturally amenable to parallel reasoning; however, the current version of Sequoia has limited support for parallel reasoning. This project aims at optimising Sequoia by identifying new and effective ways to exploit parallel reasoning. The student will perform an extensive empirical evaluation of the current version of Sequoia, identify key bottlenecks, and devise new ways to make efficient and powerful use of parallel reasoning. The student will evaluate the performance of any new versions of the system that they propose, with an eye to understanding their strengths and weaknesses. The student involved in this project will acquire intensive knowledge of stateoftheart techniques for reasoning in OWL 2. Prerequisites: Knowledge Representation and Reasoning course. Experience with concurrent programming and/or Scala is desirable 

MultiModal Partially Labelled Stream  Philip Torre  MSc  Data on large systems is often stream lined and multi modal, e.g., textual, images, videos, and or sound. All this data is being accumulated while jointly changing in distribution. Moreover, much of this data presented from the stream is only partially labelled. We seek to study the problem of training models on a partially labelled streams in multimodal setting. In particular, we seek to find new effective algorithms to performing joint selfsupervised continual learning on the unlabelled data while learning in supervised fashion the labelled portion of the stream.  
Proposal. Robust Continual Learning: Effective Learning from Adversarial Streams    Philip Torre  MSc  Continual learning is an open problem with various realworld applications. While there is a large body of work on learning from large corpus of static wellcurated data, learning from streams where the distribution of the data varies per time step is a folds more challenging problem. We seek to formulate settings where there is an adversary in the loop that can potentially alter or perturb the samples presented by the stream. The adversaries goal is break continually learning algorithms without being detected. On the other hand, efficient robust continual learners can preserve a high accuracy even under the worst case scenario where every sample presented by the stream has been adversarially manipulated. We want to formulate this new exciting problem and propose effective robust continual learning algorithms under such challenging setting.  
ARETHA: A Transparent and Respectful Virtual Assistant for the Home  Max Van Kleek  Human Centered Computing  B  C  MSc  Virtual assistants like Alexa and Siri have become hugely popular conveniences in homes worldwide. However, they also pose unprecedented privacy risks by being blackbox vehicles of corporate surveillance by the likes of Amazon, Apple and others. Is it possible to design a virtual assistant that can be guaranteed to protect user privacy, and be fully transparent and controllable by endusers? This project will involve designing and developing a fully modular, open source virtual assistant called ARETHA from the ground up, using open source components, which can provide strong guarantees about being prirvacypreserving and prouser.  
Privopticon: A privacypreserving mesh OS for selfobservation  Max Van Kleek  Human Centered Computing  B  C  MSc  Selfmonitoring has many potential applications within the home, such as the ability to understand important health and activity rhythms within a household automatically. But such monitoring activities also have associated extreme privacy risks. Can we design new kinds of sensing architectures that are designed to preserve inhabitants' privacy? We have made initial inroads of a new mesh operating system prototype for raspberry π hardware for new classes of privacypreserving selfmonitoring applications for the home, which will provide userconfigurable degrees of information fidelity, have builtin forgetting and are accountable by design. We need your help to try and evaluate different kinds of methods for achieving these goals. 

3D demos of geometric concepts  Irina Voiculescu  Computational Biology and Health Informatics  B  The Geometric Modelling course (Part B) deals with several interesting concepts which would be best visualised in a suite of applications. A coherent suite of 3D demos could easily become a useful tool for this course, as well as for users worldwide. This project would be most suitable for a candidate who already have some experience using a 3D graphics library of their choice and want to improve this skill. The mathematical concepts are welldocumented.  
3D environment for Hand Physiotherapy  Irina Voiculescu  Computational Biology and Health Informatics  B  C  After hand surgery, it is almost always necessary for patients to have physiotherapy afterwards to help with their recovery. As part of this, the patient will need to perform hand exercises at home. However, the patient may not always do the exercises correctly, or they might forget to do their exercises. The goal of this project is to use the Leap Motion to create a userfriendly GUI which a patient could use to aid them with their home exercises. The interface would show the user where their hand should be and they would then need to follow the movements. It could work from a webbased software or a downloaded software. It would need to be tailored to the patient so it contained their specific required exercises, which could be input by the physiotherapist. It would need to store data on how the patient is doing and feedback this data to the patient, and possibly also to the physiotherapist via the internet. If internetbased, patient confidentiality and security would need to be considered. This project would be performed in close collaboration with a physiotherapist, an orthopaedic hand surgeon, and a postdoctoral researcher based at the Nuffield Orthopaedic Centre.  
3D printing medical scan data  Irina Voiculescu  Computational Biology and Health Informatics  B  C  MSc  Computed tomography (CT) scanning is a ubiquitous scanning modality. It produces volumes of data representing internal parts of a human body. Scans are usually output in a standard imaging format (DICOM) and come as a series of axial slices (i.e. slices across the length of the person's body, in planes perpendicular to the imaginary straight line along the person's spine.) The slices most frequently come at a resolution of 512 x 512 voxels, achieving an accuracy of about 0.5 to 1mm of tissue per voxel, and can be viewed and analysed using a variety of tools. The distance between slices is a parameter of the scanning process and is typically much larger, about 5mm. During the analysis of CT data volumes it is often useful to correct for the large spacing between slices. For example when preparing a model for 3D printing, the axial voxels would appear elongated. These could be corrected through an interpolation process along the spinal axis. This project is about the interpolation process, either in the raw data output by the scanner, or in the postprocessed data which is being prepared for further analysis or 3D printing. The output models would ideally be files in a format compatible with 3D printing, such as STL. The main aesthetic feature of the output would be measurable as a smoothness factor, parameterisable by the user. Existing DICOM image analysis software designed within the Spatial Reasoning Group at Oxford is available to use as part of the project. 

3D stereo display of medical scan data  Irina Voiculescu, Stuart Golodetz  Computational Biology and Health Informatics, Systems  B  C  The Medical Imaging research group has been working with a variety of data sourced from CT and MRI scans. This data comes in collections of (generally greyscale) slices which together make up 3D images. Our group has developed software to generate 3D models of the major organs in these images. This project aims to develop a simple augmented reality simulation for the Oculus Rift which will render these organs within a transparent model of a human and allow the user to walk around the model so as to view the organs from any angle. This has a number of possible applications, including to train medical students and to help surgeons to explain medical procedures to their patients.  
Exact Algorithms for Complex Root Isolation  Irina Voiculescu  Computational Biology and Health Informatics  B  C  MSc  Not available in 2013/14 Isolating the complex roots of a polynomial can be achieved using subdivision algorithms. Traditional Newton methods can be applied in conjunction with interval arithmetic. Previous work (jointly with Prof Chee Yap and MSc student Narayan Kamath) has compared the performance of three operators: Moore's, Krawczyk's and HansenSengupta's. This work makes extensive use of the CORE library, which is is a collection of C++ classes for exact computation with algebraic real numbers and arbitrary precision arithmetic. CORE defines multiple levels of operation over which a program can be compiled and executed. Each of these levels provide stronger guarantees on exactness, traded against efficiency. Further extensions of this work can include (and are not limited to): (1) Extending the range of applicability of the algorithm at CORE's Level 1; (2) Making an automatic transition from CORE's Level 1 to the more detailed Level 2 when extra precision becomes necessary; (3) Designing efficiency optimisations to the current approach (such as confirming a single root or analysing areas potentially not containing a root with a view to discarding them earlier in the process); (4) Tackling the isolation problem using a continued fraction approach. The code has been included and is available within the CORE repository. Future work can continue to be carried out in consultation with Prof Yap at NYU. 

Gesture recognition using Leap Motion  Irina Voiculescu  Computational Biology and Health Informatics  B  C  MSc  Scientists in the Experimental Psychology Department study patients with a variety of motor difficulties, including apraxia  a condition usually following stroke which involves lack of control of a patient over their hands or fingers. Diagnosis and rehabilitation are traditionally carried out by Occupational Therapists. In recent years, computerbased tests have been developed in order to remove the human subjectivity from the diagnosis, and in order to enable the patient to carry out a rehabilitation programme at home. One such test involves users being asked to carry out static gestures above a Leap Motion sensor, and these gestures being scored according to a variety of criteria. A prototype has been constructed to gather data, and some data has been gathered from a few controls and patients. In order to deploy this as a clinical tool into the NHS, there is need for a systematic data collection and analysis tool, based on machine learning algorithms to help classify the data into different categories. Algorithms are also needed in order to classify data from stroke patients, and to assess the degree of severity of their apraxia. Also, the graphical user interface needs to be extended to give particular kinds of feedback to the patient in the form of home exercises, as part of a rehabilitation programme. This project was originally set up in collaboration with Prof Glyn Humphreys, Watts Professor of Experimental Psychology. Due to Glyn's untimely death a new cosupervisor needs to be found in the Experimental Psychology Department. It is unrealistic to assume this project can run in the summer of 2016. 

Identifying features in MRI scan data  Irina Voiculescu  Computational Biology and Health Informatics  B  C  MSc  In recent years, medical diagnosis using a variety of scanning modalities has become quasiuniversal and has brought about the need for computer analysis of digital scans. Members of the Spatial Reasoning research group have developed image processing software for CT (tomography) scan data. The program partitions (segments) images into regions with similar properties. These images are then analysed further so that particular features (such as bones, organs or blood vessels) can be segmented out. The team's research continues to deal with each of these two separate meanings of medical image segmentation. The existing software is written in C++ and features carefullycrafted and welldocumented data structures and algorithms for image manipulation. In certain areas of surgery (e.g. orthopaedic surgery involving hip and knee joint) the magnetic resonance scanning modality (MRI) is preferred, both because of its safety (no radiation involved) and because of its increased visualisation potential. This project is about converting MRI scan data into a format that can become compatible with existing segmentation algorithms. The data input would need to be integrated into the group's analysis software in order then to carry out 3D reconstructions and other measurements. This project is cosupervised by Professor David Murray MA, MD, FRCS (Orth), Consultant Orthopaedic Surgeon at the Nuffield Orthopaedic Centre and the Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences (NDORMS), and by Mr Hemant Pandit MBBS, MS (Orth), DNB (Orth), FRCS (Orth), DPhil (Oxon)Orthopaedic Surgeon / Honorary Senior Clinical Lecturer, Oxford Orthopaedic Engineering Centre (OOEC), NDORMS.  
Reinforcement learning techniques for games  Irina Voiculescu  Computational Biology and Health Informatics  B  C  This project is already taken for 20172018
Psychology has inspired and informed a number of machine learning methods. Decisions within an algorithm can be made so as to improve an overall aim of maximising a (cumulative) reward. Supervised learning methods in this class are known as Reinforcement Learning. A basic reinforcement learning model consists of establishing a number of environment states, a set of valid actions, and rules for transitioning between states. Applying this model to the rules of a board game means that the machine can be made to learn how to play a simple board game by playing a large number of games against itself. The goal of this project is to set up a reinforcement learning environment for a simple board game with a discrete set of states (such as Backgammon). If time permits, this will be extended to a simple geometric game (such as Pong) where the states may have to be parameterised in terms of geometric actions to be taken at each stage in the game. 

Simple drawing analysis  Irina Voiculescu  Computational Biology and Health Informatics  B  C  MSc  Scientists in the Experimental Psychology Department study patients with a variety of motor difficulties, including apraxia  a condition usually following stroke which involves lack of control of a patient over their hands or fingers. Diagnosis and rehabilitation are traditionally carried out by Occupational Therapists. In recent years, computerbased tests have been developed in order to remove the human subjectivity from the diagnosis, and in order to enable the patient to carry out a rehabilitation programme at home. One such test involves users drawing simple figures on a tablet, and these figures being scored according to a variety of criteria. Data has already been gathered from 200 or so controls, and is being analysed for a range of parameters in order to assess what a neurotypical person could achieve when drawing such simple figures. Further machine learning analysis could help classify such data into different categories. Algorithms are also needed in order to classify data from stroke patients, and to assess the degree of severity of their apraxia. This project was originally cosupervised by Prof Glyn Humphreys, Watts Professor of Experimental Psychology. Due to Glyn's untimely death a new cosupervisor needs to be found in the Experimental Psychology Department. It is unrealistic to assume this project can run in the summer of 2016. 

Efficient linear algebra for block structured matrices  Jonathan Whiteley  Computational Biology and Health Informatics  B  C  Many large systems of linear equations are sparse, i.e. if the matrix that describes the linear system is of size N times N, where N is large, there are very few nonzero entries in each row. Under these conditions there may not be enough memory to store the whole matrix, and so only the nonzero entries are stored. This prevents techniques such as LU decomposition being used to solve the linear system; instead an iterative technique such as the conjugate gradient technique for symmetric positive definite matrices, or GMRES for more general matrices is used. The number of iterations needed with these techniques can be large, rendering these techniques inefficient. To prevent this, preconditioning techniques are used  if the linear system is defined by Ax=b, then a preconditioner P is used and the system solved is instead PAx = Pb, where P is cheap to calculate and both PAx and Pb are cheap to evaluate. In this project we will investigate matrices with a block structure that arises in many fields, such as constrained optimisation and continuum mechanics. We will utilise the block structure of these matrices to heuristically derive candidate preconditioners, and compare their performances. Prerequisites: linear algebra, continuous mathematics 

Parameter recovery for models described by differential equations  Jonathan Whiteley  Computational Biology and Health Informatics  B  C  Phenomena in many fields are described by differential equations where a quantity of interest (for example, depending on the phenomena modelled, a reaction rate, a capacitance, or a subject's cardiac output) appears as a parameter in the differential equation. We then determine the quantity of interest by choosing parameters in the differential equation so that the computed solution of the differential equation using these parameters matches experimental data as closely as possible. This may be posed as an optimisation problem that may be tackled by either a classical optimisation approach (as seen in the continuous mathematics course) or a Bayesian optimisation approach. The aim of this project is to compare these approaches. Prerequisites: linear algebra, continuous mathematics, probability  
Topics in Linear Dynamical Systems  James Worrell  Automated Verification  B  C  MSc  A linear dynamical system is a discrete or continuoustime system whose dynamics is given by a linear function of the current state. Examples include Markov chains, linear recurrence sequences (such as the Fibonacci sequence), and linear differential equations. This project involves investigating the decidability and complexity of various reachability problems for linear dynamical systems. It would suit a mathematically oriented student. Linear algebra is essential, and number theory or complexity theory is desirable. A relevant paper is Ventsislav Chonev, Joël Ouaknine, James Worrell: The orbit problem in higher dimensions. STOC 2013: 941950.  
Distributed Programming with Distributed Protocols in Scala  Nobuko Yoshida  Programming Languages  C  Session type systems facilitate the development of concurrent and distributed software by enabling programmers to describe communication protocols in types. Not only does this aid understanding and correctness of communicating code, it can guarantee desirable behavioural properties (e.g. deadlockfreedom). Concurrency libraries, e.g. Effpi, enable the use of session types in realworld programming languages and systems. Unfortunately, in practice, such libraries can require a significant number of channels between protocol participants, potentially hindering maintenance and be a source of error. The goal of this project is to raise the level of abstraction in the Effpi concurrency library. It will first focus on safely abstracting away lowlevel binary channels that are currently exposed to the programmer. The project will explore possible abstractions, e.g. session endpoints and skeletons of communication. Further work would be to explore additional abstractions to improve the correctness and maintainability of concurrent software using the Effpi library. Reference: http://dx.doi.org/10.1145/3314221.3322484 

Distributed Programming with Distributed Protocols in Scala  Nobuko Yoshida  Programming Languages  B  Session type systems facilitate the development of concurrent and distributed software by enabling programmers to describe communication protocols in types. Not only does this aid understanding and correctness of communicating code, it can guarantee desirable behavioural properties (e.g. deadlockfreedom). Concurrency libraries, e.g. Effpi (Scala), enable the use of session types in realworld programming languages and systems. The goal of this project is to understand basic of session types, to implement distributed protocols in Scala and investigate its usability. Reference: [1] Session Types https://dl.acm.org/doi/pdf/10.1145/2873052 [2] http://dx.doi.org/10.1145/3314221.3322484 

Extension to Probabilistic ResourceAware Session Types  Nobuko Yoshida, Joe Paulus  Programming Languages  B  "Probabilistic session [1] types explores how uncertainty and likelihood influence communication protocols in distributed systems. We propose an extension on the session type system in which an abstract notion of success and failure can be attached either to the syntax of session types or to the both the type and process level such as explicit failure terms. The goal of which is to extend the scope of [1] showing that probabilistic analysis of binary session types [2] can be applied. Similarly insight are applied in the behavioural equivalence of probabilistically nondeterministic behaviour as well of implementations of extensions within NomosPro/PRast." [1] Probabilistic ResourceAware Session Types (acm.org) [2] Probabilistic Analysis of Binary Sessions (dagstuhl.de) Prerequisites The student wishes to learn a type theory of concurrency and communication, LCT will be helpful 

Formalising Mac Lane's comparison theorem for the EilenbergMoore constructions in Coq.  Nobuko Yoshida  Programming Languages  C  MSc  Monads serve as a means to encapsulate the impure operations within a computation. They're defined by an adjunction and, in turn, give rise to a specialized form of adjunction known as the EilenbergMoore adjunction. Mac Lane's introduction of the comparison theorem [1] provides a framework for connecting these adjunctions unified by a monad through a unique comparison functor. This theorem finds application in certain categorical contexts, as demonstrated by Jacob [2], particularly in interpreting state effects within impure programming languages. In this project, our objective is to formalize fundamental elements of category theory within the Coq proof assistant. We aim to demonstrate the value of this formalization by certifying Mac Lane's comparison theorem and Beck's monadicity theorems, showcasing the practical utility of our work. There is an existing formalization of the comparison theorem for the Kleisli constructions in Coq [3], which serves as a promising starting point. Additionally, it is worth noting that the ongoing Coq proof holds strong potential for presentation and publication at conferences such as ITP and CPP, as well as in reputable journals like JFR and JAR. [1] Saunders Mac Lane. Categories for the Working Mathematician, Graduate Texts in Mathematics, 1978. [2] Bart Jacobs. A Receipe for Stateandeffect triangles, Logical Methods in Computer Science, 2017. [3] Burak Ekici and Cezary Kaliszyk. Mac Lane’s Comparison Theorem for the Kleisli Construction Formalized in Coq, Mathematics in Computer Science, 2020. Prerequisites: a student who wishes to learn Coq is welcome. Familiarity with basic category theory. Familiarity with the Coq proof assistant is also helpful but not required 

Formalism of The Go Language  Nobuko Yoshida  Programming Languages  B  Go is a popular statically typed programming language designed and implemented by Google, and used, among others, by Uber, Netflix, and Docker. The recent release of Go 1.18 finally realised the long awaited addition of Generics (parametric polymorphism) to the Go type system. Much academic work has been conducted to ensure that generics are correctly implemented, but there is still more to do [1,2,3,4]. This project aims to give a survey of those four papers, and take one of miniprojects: (1) build some examples using the existing prototype of [1,2,3,4] or (2) survey important features included in Go 1.18 such as type inference and type sets [5] but not included in [1,2,3,4]. [1] https://dl.acm.org/doi/pdf/10.1145/3428217 [2] https://dl.acm.org/doi/10.1007/9783030890513_7 [3] https://dl.acm.org/doi/10.1007/9783031169120_7 [4] https://arxiv.org/abs/2208.06810v2 [5] https://go.googlesource.com/proposal/+/master/design/43651typeparameters.md 

Mechanisation of Distributed Protocol Specifications  Nobuko Yoshida  Programming Languages  C  Multiparty session types (MPST) offer a specification and verification framework for concurrency: communicating systems can be safely implemented in a distributed fashion, when welltyped against local session types, provided that such local types are obtained by projection of a single choreographic protocol (global type). Multiple projects are available, please get in touch with nobuko.yoshida@cs.ox.ac.uk for more detail. Possible aims are: (i) exploring and implementing an algorithms for MPST protocol compositionality or (ii) formalising correctness of advanced MPST systems (e.g., featuring merge, subtyping, or delegation). Students with a strong interest in the mechanisation of MPST in proof assistants (Coq, Isabelle, Idris) are very welcome to reach out. [1] Mechanisation Paper https://dl.acm.org/doi/10.1145/3453483.3454041 [2] Multiparty Session Types Paper https://link.springer.com/chapter/10.1007/9783030369873_5 

Modelchecking Timed Session Types  Nobuko Yoshida  Programming Languages  B  C  MSc  This project lies in the intersetion of modelchecking, verification, timed systems, and session types, It aims to tackle the critical challenge of ensuring deadlockfreedom in distributed systems employing timed session types [1, 2] and the application of timed modelchecking tools. Timed session types provide a formalism for specifying communication protocols with temporal constraints, while deadlocks pose a persistent threat to system stability. Drawing on established time modelchecking tools [3], our objective is to systematically analyse the specified models to detect and scrutinise potential deadlocks. The project also endeavours to build Timed Session Types using a bottomup approach, and then verify the typelevel properties by using a model checker, as in [4]. This research aspires to augment the reliability and correctness of concurrent and distributed systems, contributing to the progression of formal verification techniques. [1] http://mrg.doc.ic.ac.uk/publications/timedmultipartysessiontypes/CONCUR14.pdf [2] https://mrg.cs.ox.ac.uk/publications/meetingdeadlinestogether/ [3] https://uppaal.org/features/ (UPPAAL model checker) [4] https://dl.acm.org/doi/10.1145/3290343 Using a modelchecker to verify the typelevel properties for timed binary sessions, along with a literature survey on timed systems with extensions into deadlockfreedom checking could be a miniproject of shorter duration. Prerequisites: A student who wishes to learn model checking is welcome. Familiarity with the basic automata theory. Familiarity with the model checking tools will be helpful but not required. 

Probabilistic Session Types: semantics and tool development  Nobuko Yoshida, Joe Paulus  Programming Languages  C  MSc  Probabilistic session [1] types explores how uncertainty and likelihood influence communication protocols in distributed systems. Extending Probabilistic ResourceAware Session Types allows for many possibilities, they typing is based on the system of DILL [2] ( a session type system with a Curry Howard isomorphism with linear logic) and can be extended with parallel and restriction similarly alternate rules can be derived from the classical viewpoint [3] expanding typeability. This project is be a mix of both theory and practice where an extension of the implementation NomosPro/PRast. Overall, this extension aims to provide a more comprehensive framework for designing and analysing distributed systems with probabilistic and resourceaware communication protocols. The project also develops bisimulation semantics for concurrent processes. [1] Probabilistic ResourceAware Session Types (acm.org) [2] concur10.pdf (cmu.edu) [3] propositionsassessions.pdf (ed.ac.uk) Prerequisites: DPTP 

Program Transformation of Distributed Protocol Specification  Nobuko Yoshida  Programming Languages  C  Session types [1] describe patterns of communication in concurrent and distributed systems. Moreover, they guarantee certain desirable behavioural properties, such as deadlock freedom. Although session type libraries exist for a range of programming languages, there is little programmer support for their integration into existing and legacy code. This necessarily raises the level of required expertise to introduce and maintain session typed systems, and can consequently be a source of error. The goal of this project is to develop tools and techniques that can assist programmers in the introduction and manipulation of session types in legacy code. Inspired by similar work for parallelism [2], the project will focus on program transformation techniques, including refactoring [2,3,4], to develop safe and semiautomatic means to not only introduce session types, but modify existing session types in situ. [1] N. Yoshida and L. Gheri: A Very Gentle Introduction to Multiparty Session Types. ICDCIT 2020: 7393 [2] C. Brown, et al.: Refactoring GrPPI: Generic Refactoring for Generic Parallelism in C++. Int. J. Parallel Program. 48(4): 603625 (2020) [3] T. Mens and T. Tourwé: A Survey of Software Refactoring. IEEE Trans. Software Eng. 30(2): 126139 (2004) [4] S. J. Thompson and H. Li: Refactoring tools for functional languages. J. Funct. Program. 23(3): 293350 (2013) [5] R. N. S. Rowe, et al.: Characterising renaming within OCaml's module system: theory and implementation. PLDI 2019: 950965 

Rust programming language for communication and distribution.  Nobuko Yoshida  Programming Languages  C  Rust is the most beloved programming language since 2016 according to the annual survey by Stack Overflow. Thanks to its efficiency and memory safety, it is now one of the most popular languages of largescale concurrent applications such as Servo, a browser engine by Firefox, Stratis, a file system manager for Fedora, and Microsoft Azure. Memory safety is one of the core principles of Rust but does not extend to communication safety, making the implementation of deadlockfree protocols challenging. Our group has been working on implementing library for asynchronous Rust programming and we wish to tackle several challenges such as refinement types, dependent types, and reversible computing with Rust. Such projects can be both theoretical and practical. EXAMPLE (1): Dependent types and asynchronous message reordering In this project, we are interested in developing the dependent type theory for asynchronous message reordering [2] and use Rumpsteak to implement the theory [1]. EXAMPLE (2): Reversible computing In this project, we are interested in using Rust to implement a reversible process calculus. Reversible process calculi are widely studied, in particular for debugging [3]. In this project, we want to explore the usefulness of reversibility as a programming primitive, similar to [4]. Direct applications include fault tolerance (e.g. when merged with affine MPST [5]) or speculative execution. Reference: [1] Zak Cutner, Nobuko Yoshida, Martin Vassor: DeadlockFree Asynchronous Message Reordering in Rust with Multiparty Session Types. PPoPP '22 : 261  246. https://dl.acm.org/doi/10.1145/3503221.3508404 [2] Silvia Ghilezan, Jovanka Pantovic, Ivan Prokic, Alceste Scalas, Nobuko Yoshida: Precise Subtyping for Asynchronous Multiparty Sessions. POPL 2021 : 16:1  16:28. https://dl.acm.org/doi/10.1145/3434297 [3] See for instance the CauDEr: https://github.com/mistupv/cauder [4] Controlling Reversibility in HigherOrder Pi, Ivan Lanese, Claudio Antares Mezzina, Alan Schmitt & JeanBernard Stefani, https://doi.org/10.1007/9783642232176_20 [5] http://mrg.doc.ic.ac.uk/publications/affinerustprogrammingwithmultipartysessiontypes/ 

Rust programming language for communication and distribution.  Nobuko Yoshida  Programming Languages  B  Rust is the most beloved programming language since 2016 according to the annual survey by Stack Overflow. Thanks to its efficiency and memory safety, it is now one of the most popular languages of largescale concurrent applications such as Servo, a browser engine by Firefox, Stratis, a file system manager for Fedora, and Microsoft Azure. Memory safety is one of the core principles of Rust but does not extend to communication safety, making the implementation of deadlockfree protocols challenging. Our group has been working on implementing library for asynchronous Rust programming. The miniprojects focus on either: (1) survey recent work on Rust with session types and implement an example using [1]; or (2) survey recent work on Rust with session types and study a theory of asynchronous message reordering in Rust [2] Reference: [1] Zak Cutner, Nobuko Yoshida, Martin Vassor: DeadlockFree Asynchronous Message Reordering in Rust with Multiparty Session Types. PPoPP '22 : 261  246. https://dl.acm.org/doi/10.1145/3503221.3508404 [2] Silvia Ghilezan, Jovanka Pantovic, Ivan Prokic, Alceste Scalas, Nobuko Yoshida: Precise Subtyping for Asynchronous Multiparty Sessions. POPL 2021 : 16:1  16:28. https://dl.acm.org/doi/10.1145/3434297 

Survey of Mechanisation of Distributed Protocol Specifications, Session Types  Nobuko Yoshida  Programming Languages  B  Session type systems for distributed and concurrent computations, encompass concepts such as interfaces, communication protocols, and contracts. The session type of a software component specifies its expected patterns of interaction using expressive type languages, so types can be used to determine automatically whether the component interacts correctly with other components. There are several recent work on mechanisation (proof assistants) of session types (Coq, Isabelle, Idris). This project gives an updated survey on mechanisation of session types. [1] Example of mechanisation papers in Coq https://dl.acm.org/doi/10.1145/3453483.3454041 [2] Multiparty Session Types Paper https://link.springer.com/chapter/10.1007/9783030369873_5 

Survey of Session Types Literature  Nobuko Yoshida  Programming Languages  B  This project produces an updated literature survey of session types. Session type systems for distributed and concurrent computations, encompass concepts such as interfaces, communication protocols, and contracts. The session type of a software component specifies its expected patterns of interaction using expressive type languages, so types can be used to determine automatically whether the component interacts correctly with other components. Session type systems are studied in the context of theory (including automata theory, semantics, linear logic and type theories), programming languages (Java, Scala, Go, Rust, TypeScripts, PureScripts, MPIC, C, Python, Erlang, F*, F#, Haskell, OCaml, and more) and system applications. There are four surveys in past but they are already outdated. The project produces an updated survey focusing on theory, programming languages and/or applications. [1] Theory https://dl.acm.org/doi/pdf/10.1145/2873052 [2] Tools https://www.awesomebooks.com/book/9788793519824/behaviouraltypesfromtheorytotoolsriverpublishersseriesinautomationcontrolandrobotics [3] Security https://www.sciencedirect.com/science/article/pii/S2352220815000851?via%3Dihub [4] Programming Languages https://doi.org/10.1561/2500000031 

The Go Language with Generic Types  Nobuko Yoshida  Programming Languages  C  Go is a popular statically typed programming language desinged and implemented by Google, and used, among others, by Uber, Netflix, and Docker. The recent release of Go 1.18 finally realised the long awaited addition of Generics (parametric polymorphism) to the Go type system. Much academic work has been conducted to ensure that generics are correctly implemented, but there is still more to do [1,2,3,4]. This project aims to further reduce the gap between theory and practice, allowing the Go Team to improve their implementation of generics in Go. Project 1: Formalise (and prove correct) a dictionarypassing translation from Featherweight Generic Go (FGG) to a lambdacalculus with pattern matching or other suitable lowlevel language. Project 2: The current model of generics in Go (FGG) does not include important features included in Go 1.18 such as type inference and type sets [5]. The proposed project would be to formalise a true FGG (including aforementioned features) along with a correct translation. [1] https://dl.acm.org/doi/pdf/10.1145/3428217 [2] https://dl.acm.org/doi/10.1007/9783030890513_7 [3] https://dl.acm.org/doi/10.1007/9783031169120_7 [4] https://arxiv.org/abs/2208.06810v2 [5] https://go.googlesource.com/proposal/+/master/design/43651typeparameters.md 

Verifying and implementing security protocols in Rust  Nobuko Yoshida  Programming Languages  C  MSc  In this project, our goal is to develop a small working voting software like Helios [1] using Session Types. In the beginning. server will send credentials to all eligible voters before an election. On the election day, all the eligible voters will authenticate themselves and cast their ballot. At this point, the server will encrypt the ballot and will give two options to the voter, either decrypt the ballot to ensure that server is not cheating or cast it. In case of decryption, the voter will again be prompted to enter a new ballot and given two options again, decrypt or cast. A voter can decrypt the ballot as many times as possible, to ensure that server is not cheating, before casting the ballot. Once it's cast, the server will send a confirmation to the voter and post the encrypted ballot to a bulletin board. After the end of the election, the server combines all the ballots homomorphically and decrypts the final tally and declares the winner. Our first goal is to capture a simple version of Helios protocol using Session Types in NuScr (nuScribble) and generate an API, and later fill the API with Rust implementation. This project requires the knowledge of cryptography (basic understanding of ElGamal encryption), NuScibble, and Rust. [1]https://github.com/benadida/heliosserver [2]https://www.usenix.org/legacy/events/sec08/tech/full_papers/adida/adida.pdf 

Verifying security protocols (Rust)  Nobuko Yoshida  Programming Languages  B  Sigma protocols are a particularly simple and efficient kind of zeroknowledge proof and have seen wide deployment; they remain a leading kind of proof both in terms of simplicity and deployment but recent advances in succinct zeroknowledge proofs offer greater efficiency. The first efficient sigma protocol was introduced by Schnorr [2], several years before the class was defined. You can read more about sigma protocols in chapter 5 of this book [1]. To give you some insight into sigma protocols, we briefly discuss the most famous sigma protocol, the Schnorr protocol [2]. Given some public input (G, g, q, h) where G is a cyclic group of prime order q, and g and h are two generators of the group G, the prover claims that she knows a witness w for the statement h = g^w; the existence of such a w is immediate because g generates the group. However, does the prover know the witness w? In order to convince the verifier, the prover and the verifier do the following:
The verifier accepts if g^t = c ∗ h^e, otherwise rejects. In this project, the goal is to use session types to capture the communication between the Prove and the Verifier. It is a very simple binary session type with three messages. Our first step will be to encode the Schnorr protocol, described above, in NuScribble, generate Rust API fron NuScribble encoding, and fill the rest of the cryptographic code. Later, we will develop the Parallel composition, AND composition, OR composition, and NEQ composition of sigma protocols (see the chapter 5 of [1]) [1] https://www.win.tue.nl/~berry/2WC13/LectureNotes.pdf [2] https://link.springer.com/article/10.1007/BF00196725 

Ethical Web and Data Architecture for Gig Workers  Jun Zhao, Rui Zhao  Human Centered Computing  MSc  Description: The Oxford Ethical Web and Data Architecture (EWADA) project (ewada.ox.ac.uk) explores how to redecentralise the power on the web and provide individuals with more data autonomy and transparency. EWADA works closely with the Solid project (solidproject.org), an opensource project started in MIT, to investigate how to build such an ethical web for everyone. One area we are looking into is to support Gig Workers (such as Uber driver) gaining more transparency about how their pay rate is calculated by large platforms and how they can gain more autonomy. Prerequisites: In this project, you will explore what tools and technologies are already out there to support gig workers free from the monopoly of large platforms. You will then design and implement a Solidbased prototype to give gig workers more control of their data and allow them to gain value from their data without losing control of data privacy. You will evaluate the project by assessing the robustness and soundness of the prototype. For able students, you could also work with Gig workers to assess the usefulness of the system.  
Fairness in Distributed/Decentralised Machine Learning  Jun Zhao, Naman Goel  Human Centered Computing  MSc  An emerging paradigm of machine learning in distributed settings is federated learning. The idea is that data of different agents stays with them and learning happens in a relatively "decentralised" manner. One notable aspect of federated learning is the noniid nature of the data across different clients. In the first part of the project, the student will identify the technical challenges that this poses for fairness properties of the trained machine learning model. The goal of fair machine learning is to ensure that decisions taken by machine learning systems don't discriminate based on sensitive attributes like race and gender. This part will involve getting familiar with the literature and some experiments or simulations. The second part of the project will involve proposing, implementing and evaluating solutions to address the problem. Further, the student will also have the opportunity to work on proving formal guarantees for their algorithm. Prerequisites: Good understanding and handson experience with machine learning and statistics, proficiency in Python, interest in ML fairness. Students are encouraged to reach out to Naman Goel (naman.goel@cs.ox.ac.uk) to discuss more project ideas related to above topics.  
Fairness in Dynamic Settings  Jun Zhao, Naman Goel  Human Centered Computing  MSc  The goal of fair machine learning is to ensure that decisions taken by machine learning systems don't discriminate based on sensitive attributes like race and gender. Most of the fair machine learning literature focuses on settings in which individuals can't change their features that a model takes as input. However, in many situations, the individuals can modify their features to get a positive decision based on the apriori information or feedback/explanation provided by the model/decisionmaker. How should fairness be defined in these settings, such that certain individuals or groups do not face disproportionate obstacles in achieving better outcomes? In this project, the student will start by empirically evaluating the existing approaches to fairness. In collaboration with the supervisor, the student will then define the appropriate notion(s) of fairness for the above settings and develop techniques to satisfy those notions. The student will implement the technique and use both simulations and realworld datasets to evaluate the effectiveness of the technique. Further, depending on interest and time, there will also be the opportunity to work on proving formal guarantees for the algorithm and/or a human subject experiment. Prerequisites: Familiarity with machine learning, proficiency in Python, interest in ML fairness . Students are encouraged to reach out to Naman Goel (naman.goel@cs.ox.ac.uk) to discuss more project ideas related to above topics.  
Fairness in HumanAI Collaboration  Jun Zhao, Naman Goel  Human Centered Computing  MSc  The goal of fair machine learning is to ensure that decisions taken by machine learning systems don't discriminate based on sensitive attributes like race and gender. A limitation/weakness of current approaches to fair machine learning is that they do not account for the behaviour of the humans who use these ML systems as decisionsupport tools. This may result in fairness being violated in the final decisions. In this project, the student will explore this aspect more formally. In particular, we would be interested in understanding under what conditions, fairness outcomes can improve if the human has more control over the output and internal design of the ML system. The student will be encouraged to use simulations and/or analytical methods to derive the conditions. Depending on time and interest, there will also be an opportunity to conduct human subject experiments. Prerequisites: Familiarity with machine learning, proficiency in Python, interest in ml fairness and behavioural studies. Students are encouraged to reach out to Naman Goel (naman.goel@cs.ox.ac.uk) to discuss more project ideas related to above topics.  
Improving the experience of MPC in decentralized contexts  Jun Zhao, Rui Zhao  Human Centered Computing  C  MSc  SoLiD (Social Linked Data) or other Personal Data Storage (PDS) systems. Such decentralized PDS is gaining attention in recent years, and supporting MPC in such scenarios is crucial for privacy and user control. We have benchmarked two different models of applying MPC in decentralized contexts, and developed an architecture implementation to perform MPC on SoLiD, a popular framework for enabling decentralised data autonomy for web users. This implementation demonstrates the architecture well, but is not easy to use by normal users. This project aims to investigate a more usable privacypreserving computation for the users by optimising the efficiency of the computation or preconfigure MPC parameters (protocol, number of clients, number of players, etc) for the users.  
User Interface for understanding and reasoning on data Terms of Use  Jun Zhao, Rui Zhao  Human Centered Computing  C  MSc  Description Ticking without reading “I have read and agree to the Terms of Service” when registering online accounts is known as “the biggest lie on the Internet”. We are all guilty of ticking without reading it, but we recognize the necessity of the existence of these terms. Data Terms of Use (DToU) is a similar but more approachable concept. It specifies what the data users need to follow before, during, and after using the data. They tend to be humanunderstandable, to promote more data users comply with them. In a decentralised web context  where everyone becomes a data provider, everyone will need to write their own DToU. In addition, not only do they need to govern the direct use of data, the use of derived data is also subject to such DToU. Handling such DToU is often labourconsuming and errorprone. Formalisation and automated reasoning have the potential to improve the handling and understanding of such DToU. We have previously developed a formal DToU language and the reasoning engine for decentralised contexts, with potential to be used in wider scenarios. In this project, the student is expected to investigate how usable this formal logicbased DToU language is and assess the performance of the framework in the decentralised setting at a large scale. 

Topics in Algorithms, Complexity, and Combinatorial Optimisation  Standa Živný  Algorithms and Complexity Theory  C  MSc  Prof Zivny is willing to supervise in the area of algorithms, complexity, and combinatorial optimisation. In particular, on problems related to convex relaxations (linear and semidefinite programming relaxations), submodular functions, and algorithms for and complexity of homomorphisms problems and Constraint Satisfaction Problems. Examples of supervised projects involve extensions of mincuts in graphs, analysis of randomised algorithms for graph and hypergraph colourings, and sparsification of graphs and hypergraphs. The projects would suit mathematically oriented students, with interest in rigorous analysis of algorithms and applications of combinatorics and probabilistic methods to computer science. 