AI only needs to listen to the sound of keystrokes to predict the content, achieving an accuracy rate of up to 95%

A group of researchers from Cornell University (Ithaca, New York) has utilized AI to accurately predict inputted content solely by listening to the sound of keystrokes.

Accordingly, the research team used a database of keyboard typing sounds to train the Artificial Intelligence. Through this method, the AI can predict the content being typed on the keyboard with an accuracy of up to 95%. This accuracy only drops to 93% when using Zoom to train the system.

However, the training approach of the research team from Cornell University still has several limitations. For example, the AI’s predictive ability cannot be immediately applied to any random type of keyboard.

In other words, the AI needs some time to ‘familiarize’ itself with each specific type of keyboard, as each keyboard typing sound is used as a reference for the AI to predict characters during the training process. The actual training of the AI model can be conducted on the spot using a microphone, or even remotely through applications like Zoom to record the keyboard typing sounds.

It is known that the research team utilized a MacBook Pro to demonstrate the AI’s capability in predicting content. They pressed 36 different keys, with each key being pressed 25 times. This formed the basis for the AI model to recognize which character was being typed, corresponding to the sound of keyboard typing. Despite the minor variations in the audio waveform produced for each keystroke, the AI accurately identified each key astonishingly.

Of course, employing the AI model for malicious purposes (such as data theft) is not a sound idea, as there still exist numerous vulnerabilities.

Certainly, there have been concerns about the potential misuse of the AI model for nefarious purposes (like data theft). Nevertheless, this type of attack is not without weaknesses, according to the research team.

For example, simply altering the typing style is enough to diminish the accuracy of the AI’s content prediction. Using a gentle touch typing style, the AI’s accuracy in recognizing content drops from 64% to 40%. Additionally, users can employ software to introduce noise as input interference (using white noise) or introduce extra keystrokes to ‘confuse’ the AI.

Different keyboard types themselves also bring about varying risks in terms of data theft by AI. For instance, the accuracy of AI’s data prediction tends to be higher when users utilize mechanical keyboards, which generate louder typing sounds compared to regular keyboards. However, even when using membrane keyboards (which use a rubber dome instead of individual switches under each key), enough sound is produced to train the AI model.

As a result, the most effective way to mitigate this type of attack is to implement a software-based solution rather than switching from a ‘noisy’ keyboard to a quieter one.

According to Tomshardware