Anti-Spoofing Voice Commands: A Generic Wireless Assisted Design

This paper presents an anti-spoofing design to verify whether a voice command is spoken by one live legal user, which supplements existing speech recognition systems and could enable new application potentials when many crucial voice commands need a higher-standard verification in applications. In the literature, verifying the liveness and legality of the command's speaker has been studied separately. However, to accept a voice command from a live legal user, prior solutions cannot be combined directly due to two reasons. First, previous methods have introduced various sensing channels for the liveness detection, while the safety of a sensing channel itself cannot be guaranteed. Second, a direct combination is also vulnerable when an attacker plays a recorded voice command from the legal user and mimics this user to speak the command simultaneously. In this paper, we introduce an anti-spoofing sensing channel to fulfill the design. More importantly, our design provides a generic interface to form the sensing channel, which is compatible to a variety of widely-used signals, including RFID, Wi-Fi and acoustic signals. This offers a flexibility to balance the system cost and verification requirement. We develop a prototype system with three versions by using these sensing signals. We conduct extensive experiments in six different real-world environments under a variety of settings to examine the effectiveness of our design.