According to Bleeping Computer, a research team from universities in the UK has trained a deep learning model that is very surprising and also alarming. The model can play the data by listening to the keyboard strokes.
In other words, the model can listen to your keyboard strokes with the help of a microphone and steal your private information such as passwords. You can access the research article published as a PDF by clicking the link here.
The model can listen with 95% accuracy.
Even more surprising is that the model can make 95% accurate predictions via a microphone placed next to the keyboard. It is stated that a success rate of up to 93% is achieved when keystrokes are recorded over a video conferencing application such as Zoom.
As if the current cyber threats weren’t enough, this type of attack, which brings a new concern into our lives, shows that not only passwords but also other private information such as messages can be stolen. For such an attack, the scientists say, attackers would need to record keystrokes using a nearby microphone or a malware-infected smartphone. In addition, it is added that applications such as Zoom can also be risky, as seen in the experiments.
The researchers trained the model by pressing each of the 36 keys 25 times on a MacBook Pro and recording the sounds of each stroke with the help of the iPhone 13 mini. They then produced waveforms and spectrograms that visualized the differences for each key using the registers and used these to train an image classifier called CoAtNet.
In the experiments, they put the iPhone 13 Mini 17 cm away from the computer. In the end, CoAtNet achieved the 95% and 93% accuracy we shared above. In the experiment in Skype, there was a rate of 91.7%.
Scientists state that using a different spelling or random passwords can be used as a precaution against such attacks.
51411
