Stefan Mileschin

While they sounded silly before

A team of researchers from Durham University, University of Surrey and Royal Holloway University of London, has trained a deep learning model that can steal data from keyboard keystrokes recorded using a microphone with an accuracy of 95 per cent.

Using high-powered microphones to listen to a person’s keystrokes has been possible for a while, but it has been seen as “too silly” or “too James Bond” to be practical.

Acoustic attacks have become much simpler due to the abundance of microphone-bearing devices that can achieve high-quality audio captures. Still, it seems that by adding machine learning, sound-based side-channel attacks are feasible and a lot more dangerous than previously anticipated.

The boffins achieved 95 per cent accuracy from the smartphone recordings, 93 per cent from Zoom recordings, and 91.7 per cent from Skype.

A possible defence around the attack is to include white noise, "software-based keystroke audio filters," switching to password managers — and using biometric authentication.

“Our results prove the practicality of these side channel attacks via off-the-shelf equipment and algorithms. We discuss a series of mitigation methods to protect users against these attacks,” the boffins said.

https://fudzilla.com/news/ai/57372-a...better-with-ai