Sniffing Location Privacy of Video Conference Users Using Free Audio Channels

Since the outbreak of the COVID-19 pandemic, video conferencing apps have been more broadly used to connect geographically distant people for work, school, and social interactions. These apps simulate “in-person” meetings with streamed audio and provide users with full control of their privacy. For instance, users can conveniently disable their microphones whenever they feel the need for privacy following common senses: 1) Audio signals containing semantic or contextual information pose privacy concerns; 2) Microphones are relevant only to acoustic privacy; 3) Meeting participants cannot actively intrude on each other's privacy but only opportunistically exploit accidental privacy leakages or mistakes. This paper investigates the privacy leakages that defy these assumptions. We find that any meeting participant can actively and covertly probe others' location privacy even when the webcam is disabled or virtual backgrounds are used to hide locations. More specifically, the legitimate two-way audio channel of video conferencing facilitates remote acoustic sensing, allowing an attacker to probe the users' physical surroundings and receive location-specific echo signals. However, all video conferencing systems utilize echo cancellation functions to prevent audio feedback, which inherently stops active sensing. To address this challenge, we develop a transformer-based algorithm and leverage the encoders of generative AI to counteract echo cancellation and extract stable location embeddings from severely distorted echo sounds. Furthermore, we propose two types of active acoustic sensing attacks: the in-channel echo attack, which breaks through echo cancellation by using carefully crafted signals, and the off-channel echo attack, which exploits third-party media sounds (e.g., email notification tones) to evade cancellation. We test these attacks on commercial video conferencing apps, such as Zoom, Teams, and Skype. When using only a single probing sound, our methods achieve 88.3% accuracy in recognizing recurrent places and 88.5% accuracy in identifying the contexts of new (unseen or untagged) places.