AI Deciphers Passwords from Keyboard Sounds

On August 3, 2023, three British researchers published a scientific article on an AI that can identify the letters typed on a keyboard based on their sound. The team developed a deep-learning algorithm named CoAtNet, which is fed by spectrograms representing the sound each key produces. The researchers then trained it to recognize passwords from these sounds.

CoAtNet put up an impressive performance, achieving a 95% deciphering rate with an iPhone placed twenty centimeters away from a 2021 MacBook. However, the technology presents two major limitations.

Firstly, CoAtNet must be adapted to each make of the keyboard, as the sounds differ from one keyboard to another. The second constraint lies in distinguishing between the various purposes of the Shift key: the AI can identify the press but has more trouble detecting the release. Capitalizing letters at random in a password, which is a recommended practice in digital hygiene, can throw off the algorithm.

Nonetheless, the study highlights the need to address the potential threat of attacks that exploit sounds emitted by devices.