Exploiting Playback Device's Effect on Multi-channel Audio to Secure Voice Assistants.

Voice Assistant Devices (VADs) such as Alexa, Google Now, and Siri have become increasingly popular because of their various voice-enabled features, including online shopping, controlling smart home appliances, accessing banking services, and more. However, it also brings unique security issues like voice replay attacks, where an attacker can generate a malicious voice command via a compromised playback device near the VAD. The usage of VADs in users' daily tasks makes such attack detection more important. To detect such attacks, we propose a defense system leveraging the impact of playback devices on the bass frequency region (0 to 500 Hz) of multi-channel audios. Currently, no prior work has exploited the playback device's impact on the bass frequency region of the multi-channel audios to prevent voice replay attacks. Specifically, our system divides the bass region of each channel into ten sub-bands and computes the total percentage of the overall signal's power presented in each sub-band. To make the system more robust against advanced audio attacks, we also extract the Modified Group Delay Function (MODGDF) cepstral coefficients of the bass area for phase features. Then, our proposed system applies a support vector machine (SVM) classifier to infer whether a human or a compromised playback device initiates the voice command. The system is then tested against a public multi-channel replay attack dataset. The system performance is checked in four different environmental conditions, achieving a maximum of 1.21% Equal Error Rate (EER). Our experimental results also show that incorporating more audio channels improves the attack detection performance.