Supervisor

Suitable for

Abstract

Co-supervised 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 pre-processing 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.

Pre-requisites: Knowledge of data analysis.