AI eavesdrops on typing: introduced side-channel attack on keystrokes

A team of researchers recorded keystrokes on a laptop using a conventional smartphone microphone and used it to train an AI model using deep learning. Result: The artificial intelligence then correctly recorded the keystrokes on the laptop with an accuracy of 95 percent. And this even works without sound recording on site, only via a conference via zoom.

Advertisement

Keys detected at the stop

Joshua Harrison, Ehsan Toreini and Maryam Mehrnezhad from the Universities of Durham and Surrey and the Royal Holloway University of London present their project in a scientific article (PDF file). In it, they point out that as deep learning continues to improve and as microphones and online connections become ubiquitous, the threat of side-channel attacks via keyboards increases.

In their test arrangement, the iPhone was placed a few centimeters next to a MacBook Pro and recorded the sound of the keystrokes. After training their AI model (the image-focused CoAtNet) on all MacBook Pro keystrokes (each key was typed 25 times), the smartphone microphone’s keystroke detection rate was 95 percent. If they used a zoom conference for training, the rate was still an impressive 93 percent.

Processing the recorded keystrokes

(Image: arxiv.org, Harrison/Toreini/Multinezhad)

Stops visualized for image comparison

From the recordings, they generated waveforms and spectrograms, which allowed them to visualize the differences for each key. Using the spectrogram images, they trained their AI model and, after some experimentation with numerous parameters, achieved the highest level of accuracy in determining a keystroke based on its tone.

The project opens up the worrying possibility that keystrokes on one’s own computer could be spied out at random within a short period of time, and passwords and actually confidential messages could be overheard. As the experiment shows, even a very quiet keyboard (like the one used on the MacBook Pro from 2021) is vulnerable. So if you tend to be a bit paranoid about this, you should reconsider using background noise while typing.

Advertisement

The side channel attack by the three researchers is vaguely reminiscent of an attack on keystrokes via sound recording and thermal imaging camera from 2019, but radically simplifies the procedure and opens up comparatively huge potential for such an attack.

(tiw)

Go to home page
#eavesdrops #typing #introduced #sidechannel #attack #keystrokes