QIF-T0083

Acoustic keystroke inference (typing sound classification for credential extraction)

Each key on a keyboard produces a subtly different acoustic signature based on its position, the mechanical structure beneath it, and the user's typing dynamics. Harrison & Matyunin (2023) achieved 95% keystroke classification accuracy using a deep learning model trained on laptop keyboard audio captured by a nearby phone. The attack works over Zoom/Skype calls (Compagno et al. 2017), enabling remote credential theft during video conferences. Attack scenarios: (1) malicious app with microphone access on the same desk, (2) nearby compromised smart speaker, (3) during a video call where keyboard sounds leak through the microphone. This technique pairs with accelerometer keystroke inference (T0087) for multi-modal confirmation, significantly reducing error rates.