High-Performance Optical Sensors in Mice expose a critical vulnerability — one where confidential user speech can be leaked.

Attackers can exploit these sensors’ ever-increasing polling rate and sensitivity to emulate a makeshift microphone and covertly eavesdrop on unsuspecting users.

We present an attack vector that capitalizes on acoustic vibrations propagated through the user’s work surface, and we show that existing consumer-grade mice can detect these vibrations. However, the collected signal is low-quality and suffers from non-uniform sampling, a non-linear frequency response, and extreme quantization.

We introduce Mic-E-Mouse, a pipeline consisting of successive signal processing and machine learning techniques to overcome these challenges and achieve intelligible reconstruction of user speech. We measure Mic-E-Mouse against consumer-grade sensors on the VCTK and AudioMNIST speech datasets, and we achieve an SI-SNR increase of +19𝑑𝐵, a Speaker-Recognition accuracy of 80% on the automated tests and a WER of 16.79% on the human study.