Overview

• The research shows high‑DPI, low‑latency optical sensors can pick up desk vibrations that are processed into intelligible audio using filtering and transformer‑based models.
• Experiments reported up to a +19 dB signal improvement with speech recognition accuracy ranging from roughly 42% to 61% on common datasets, using hardware as inexpensive as a $35 mouse.
• Attackers need only access to normal mouse movement packets, which many applications can read, with data exfiltration and audio reconstruction performed remotely and often unnoticed by users.
• Effectiveness depends on conditions: flat, uncovered surfaces and quiet rooms improve capture, while mouse mats, desk covers, and noisier environments hinder it.
• UC Irvine posted the paper to arXiv on September 16 and a demo video, and researchers note the approach may detect other vibrations such as footsteps and potentially keystrokes under some conditions.