Deep Learning-Based Acoustic Side Channel Attack on Keyboards

Harrison, Toreini and Mehrnezhad present a practical acoustic side channel attack that uses a deep learning model to classify laptop keystrokes from sound captured by a smartphone microphone. The authors report 95% accuracy when the classifier is trained on keystrokes recorded by a nearby phone, which they describe as the highest accuracy seen without a language model. Training on keystrokes recorded through the video-conferencing software Zoom yielded 93% accuracy, which the paper describes as a new best for that medium. The authors conclude that the attacks are practical using off-the-shelf equipment and algorithms, and they discuss mitigation methods.