Acoustic-channel attack and defence methods for personal voice assistants

Cheng, Peng (2020) Acoustic-channel attack and defence methods for personal voice assistants. PhD thesis, UNSPECIFIED.

Text (2020PengPhD)
2020PengPhD.pdf - Published Version
Download (19MB)

Abstract

Personal Voice Assistants (PVAs) are increasingly used as interface to digital environments. Voice commands are used to interact with phones, smart homes or cars. In the US alone the number of smart speakers such as Amazon’s Echo and Google Home has grown by 78% to 118.5 million and 21% of the US population own at least one device. Given the increasing dependency of society on PVAs, security and privacy of these has become a major concern of users, manufacturers and policy makers. Consequently, a steep increase in research efforts addressing security and privacy of PVAs can be observed in recent years. While some security and privacy research applicable to the PVA domain predates their recent increase in popularity and many new research strands have emerged, there lacks research dedicated to PVA security and privacy. The most important interaction interface between users and a PVA is the acoustic channel and acoustic channel related security and privacy studies are desirable and required. The aim of the work presented in this thesis is to enhance the cognition of security and privacy issues of PVA usage related to the acoustic channel, to propose principles and solutions to key usage scenarios to mitigate potential security threats, and to present a novel type of dangerous attack which can be launched only by using a PVA alone. The five core contributions of this thesis are: (i) a taxonomy is built for the research domain of PVA security and privacy issues related to acoustic channel. An extensive research overview on the state of the art is provided, describing a comprehensive research map for PVA security and privacy. It is also shown in this taxonomy where the contributions of this thesis lie; (ii) Work has emerged aiming to generate adversarial audio inputs which sound harmless to humans but can trick a PVA to recognise harmful commands. The majority of work has been focused on the attack side, but there rarely exists work on how to defend against this type of attack. A defence method against white-box adversarial commands is proposed and implemented as a prototype. It is shown that a defence Automatic Speech Recognition (ASR) can work in parallel with the PVA’s main one, and adversarial audio input is detected if the difference in the speech decoding results between both ASR surpasses a threshold. It is demonstrated that an ASR that differs in architecture and/or training data from the the PVA’s main ASR is usable as protection ASR; (iii) PVAs continuously monitor conversations which may be transported to a cloud back end where they are stored, processed and maybe even passed on to other service providers. A user has limited control over this process when a PVA is triggered without user’s intent or a PVA belongs to others. A user is unable to control the recording behaviour of surrounding PVAs, unable to signal privacy requirements and unable to track conversation recordings. An acoustic tagging solution is proposed aiming to embed additional information into acoustic signals processed by PVAs. A user employs a tagging device which emits an acoustic signal when PVA activity is assumed. Any active PVA will embed this tag into their recorded audio stream. The tag may signal a cooperating PVA or back-end system that a user has not given a recording consent. The tag may also be used to trace when and where a recording was taken if necessary. A prototype tagging device based on PocketSphinx is implemented. Using Google Home Mini as the PVA, it is demonstrated that the device can tag conversations and the tagging signal can be retrieved from conversations stored in the Google back-end system; (iv) Acoustic tagging provides users the capability to signal their permission to the back-end PVA service, and another solution inspired by Denial of Service (DoS) is proposed as well for protecting user privacy. Although PVAs are very helpful, they are also continuously monitoring conversations. When a PVA detects a wake word, the immediately following conversation is recorded and transported to a cloud system for further analysis. An active protection mechanism is proposed: reactive jamming. A Protection Jamming Device (PJD) is employed to observe conversations. Upon detection of a PVA wake word the PJD emits an acoustic jamming signal. The PJD must detect the wake word faster than the PVA such that the jamming signal still prevents wake word detection by the PVA. An evaluation of the effectiveness of different jamming signals and overlap between wake words and the jamming signals is carried out. 100% jamming success can be achieved with an overlap of at least 60% with a negligible false positive rate; (v) Acoustic components (speakers and microphones) on a PVA can potentially be re-purposed to achieve acoustic sensing. This has great security and privacy implication due to the key role of PVAs in digital environments. The first active acoustic side-channel attack is proposed. Speakers are used to emit human inaudible acoustic signals and the echo is recorded via microphones, turning the acoustic system of a smartphone into a sonar system. The echo signal can be used to profile user interaction with the device. For example, a victim’s finger movement can be monitored to steal Android unlock patterns. The number of candidate unlock patterns that an attacker must try to authenticate herself to a Samsung S4 phone can be reduced by up to 70% using this novel unnoticeable acoustic side-channel.