Keystroke data compromised in novel acoustic side-channel attack

Keyboard keystroke data has been exfiltrated with 95% accuracy in a novel side-channel attack technique involving a deep learning model trained using a smartphone-integrated microphone, BleepingComputer reports. Meanwhile, the model dubbed "CoAtNet," which was trained using spectrogram images, yielded 93% and 91.7% accuracy in obtaining keystroke data recorded through Zoom and Skype, respectively, according to a study by researchers from Durham University, University of Surrey, and Royal Holloway University of London. The findings indicate the growing threat of sound-based side-channel attacks amid the increasing prevalence of microphone-containing devices and machine learning advancements. Such a threat could be mitigated through modifications in typing styles or the usage of randomized passwords, said researchers, who also suggested the utilization of software-based keystroke audio filters and other software tools for white noise and keystroke sound reproduction. Users have also been urged to leverage biometric authentication and password managers to reduce the risk of such attacks.