The neural network was taught to steal passwords using the sounds of the keys being pressed

A group of scientists from British universities trained AI to guess the sounds of the keys being pressed and thus determine the entered passwords. Recording is done using the microphone in the device or a nearby smartphone.

Tests have shown that the neural network can determine the pressed letters, numbers and symbols with an accuracy of up to 95%.

The training was conducted on an Apple MacBook Pro laptop, in which, as it turned out, all the keys have a unique sound. To memorize them, the neural network took 25 listenings (but with signal-enhancing processing). That is, if necessary, AI can quickly be taught to detect other keyboards.

The experiments also tested the capture of sound through the iPhone 13 mini, located 17 centimeters from the target, Zoom and Skype. The accuracy of guessing on a smartphone remained at the level of 95%, and through video conferencing services it slightly decreased to 93% and 91.7%, respectively.

The scientists note that this attack model was effective even with very quiet keyboards. To protect against the acoustic method of stealing passwords, they recommend using biometric methods. authentication, where possible, or password managers, which also eliminate the need to enter a code manually.