RF-Mic: Live Voice Eavesdropping via Capturing Subtle Facial Speech Dynamics Leveraging RFID

Eavesdropping on human voice is one of the most common but harmful threats to personal privacy. Glasses are in direct contact with human face, which could sense facial motions when users speak, so human speech contents could be inferred by sensing the movements of glasses. In this paper, we present a live voice eavesdropping method, RF-Mic, which utilizes common glasses attached with a low-cost RFID tag to sense subtle facial speech dynamics for inferring possible voice contents. When a user with a glasses, which is attached an RFID tag on the glass bridge, is speaking, RF-Mic first collects RF signals through forward propagation and backscattering. Then, body motion interference is eliminated from the collected RF signals through a proposed Conditional Denoising AutoEncoder (CDAE) network. Next, RF-Mic extracts three kinds of facial speech dynamic features (i.e., facial movements, bone-borne vibrations, and airborne vibrations) by designing three different deep-learning models. Based on the extracted features, a facial speech dynamics model is constructed for live voice eavesdropping. Extensive experiments in different real environments demonstrate that RF-Mic can achieve robust and accurate human live voice eavesdropping.