The Security of the Speech Interface: A Modelling Framework and Proposals for New Defence Mechanisms
Mary K. Bispham‚ Ioannis Agrafiotis and Michael Goldsmith
This paper presents an attack and defence modelling framework for conceptualising the security of the speech interface. The modelling framework is based on the Observe-Orient-Decide-Act (OODA) loop model, which has been used to analyse adversarial interactions in a number of other areas. We map the different types of attacks that may be executed via the speech interface to the modelling framework, and present a critical analysis of the currently available defences for countering such attacks, with reference to the modelling framework. The paper then presents proposals for the development of new defence mechanisms that are grounded in the critical analysis of current defences. These proposals envisage a defence capability that would enable voice-controlled systems to detect potential attacks as part of their dialogue management functionality. In accordance with this high-level defence concept, the paper presents two specific proposals for defence mechanisms to be implemented as part of dialogue management functionality to counter attacks that exploit unintended functionality in speech recognition functionality and natural language understanding functionality. These defence mechanisms are based on the novel application of two existing technologies for security purposes. The specific proposals include the results of two feasibility tests that investigate the effectiveness of the proposed mechanisms in defending against the relevant type of attack.