This AI Model Can Steal Data by Listening to Keystrokes with 95% Accuracy

You need to be aware of many aspects when trying to protect your data from those who want it. While there are many ways to do it, researchers from British universities believe we overlook an acoustic side of malware attacks.

They trained an AI model to listen to your keystrokes and thus steal your data with 95% accuracy. They used this deep learning method on a Zoom call and reported an accuracy of 93% – it's better (for users) but still very high and dangerous.

To set up the experiment, the authors used an iPhone microphone and then trained the AI on keystrokes, so it looks shockingly easy to steal passwords and classified information even without special hardware.

The researchers used a MacBook Pro and the iPhone 13 mini placed 17cm away from it on a folded piece of micro-fiber cloth to remove desk vibration. As for the Zoom call, they recorded keystrokes using the built-in function of the software. "Recording in this manner required no access to the victim’s environment and in this case, did not require any infiltration of their device or connection."

It is noted that while quieter keyboards are harder to decipher, it is still possible, so the authors suggest changing your typing style, for example turning to touch typing, or using randomized passwords featuring multiple cases.

You could also play sounds over a speaker near the broadcasting microphone and mix sounds into the transmitted audio locally. The best solution is adding randomly generated fake keystrokes, according to the paper, and a secondary device or biometric check to allow access to data.

This research shows another creative way someone can steal data and hopefully aids in creating new countermeasures. 

If you're interested in the results, read about them here and join our 80 Level Talent platform and our Telegram channel, follow us on Threads, Instagram, Twitter, and LinkedIn, where we share breakdowns, the latest news, awesome artworks, and more.