Protecting Voice-Controllable Devices Against Self-Issued Voice Commands

Self-issued voice commands leverage the voice-controllable device’s internal speaker to issue malicious voice commands to the device itself. These attacks are a class of voice spoofing attacks particularly challenging to protect from, as it is very hard for a countermeasure solution to infer whether the command comes from an external entity or from the device itself. In this paper, we propose a countermeasure against self-issued voice commands by training a Twin Neural Network to recognise the differences between what is being played and what is being recorded by the voice-controllable device. In fact, these audios are very similar in case of voice command self-issue attacks and different in case of legitimate commands. We start with a security and usability trade-off analysis of countermeasures against voice spoofing attacks, by describing different classes of synthesised voice commands that need to be blocked or allowed, depending on the necessities of the user. Then, we present our solution to protect voice-controllable devices from self-issued commands and show that it correctly classifies commands in the benign (real-user) and malign (self-issued) categories 97% of the times on average. We compare this result with state-of-the-art anomaly detection techniques as a baseline and show that our solution outperforms them. Furthermore, we instantiate our countermeasure on three different classes of devices to measure its performance, and we find that the additional overhead is negligible. Finally, we measure the usability impact of our solution when users interact with the tested device under different conditions, showing that our solution is resistant to environmental changes and regardless of the identity of the user issuing the commands.