Emerging Sensor-Based Attacks and Defenses Against Audio Deepfakes and Eavesdropping

Researchers at Texas Tech University have developed a prototype system that enhances voice authentication by combining traditional audio verification with physical measurements of jaw and cheek movements. This inertial-speech verification approach uses helmet-mounted motion sensors to create a unique motion profile for each speaker, making it significantly harder for attackers to impersonate someone using only voice deepfakes. The system addresses the growing threat of generative AI-powered voice cloning, which has already been used in scams and fraudulent approvals, by requiring attackers to also replicate the physical act of speaking, not just the sound.

Separately, researchers at the University of California, Irvine, have demonstrated a theoretical attack called "Mic-E-Mouse," which leverages the high sensitivity of modern optical mouse sensors to capture minute vibrations caused by nearby conversations. Under specific conditions—using mice with extremely high DPI and polling rates—adversaries could potentially eavesdrop on speech in secure environments by intercepting mouse data. Both studies highlight the evolving risks and countermeasures in the realm of audio-based attacks, emphasizing the need for multi-layered defenses as sensor technology and generative AI continue to advance.