The researchers successfully trained the deep learning model to steal data through keystroke, with an accuracy of 95%.

CTOnews.com August 7, did you know that what you typed on the keyboard might be overheard? Some researchers in the UK have developed a deep learning model that can capture and decode keystrokes to steal sensitive information such as usernames, passwords and messages with 95 per cent accuracy.

Image source PexelsCTOnews.com notes that this voice recognition algorithm can monitor users' keystrokes through videoconferencing software such as Zoom and Skype without accessing the device microphone, and infer what they enter, but the accuracy will drop to 93% and 91.7%, respectively.

The study reveals how deep learning can be used to develop new types of malware, using sound to steal information such as credit card numbers, messages, conversations and other personal information. With the progress of machine learning and the popularity of cheap and high-quality microphones in the market, voice-based attacks are more feasible than other methods limited by data transmission speed and distance.

How does this voice recognition algorithm work? Using a MacBook Pro laptop, the researchers tapped each of the 36 keys 25 times and recorded the sound produced by each key, which was recorded using iPhone 13 mini, 17 centimeters away from the laptop.

Based on the recording, the researchers generated a waveform and a spectrogram that distinguished each key. Then, use the unique sound of each key to train an image classifier called "CoAtNet", which can predict which key is pressed on the keyboard.

According to research papers, users can protect themselves from such attacks by changing their typing patterns or using complex random passwords. White noise or software that mimics keystroke can also be used to reduce the accuracy of the model.

Currently, the best way to deal with such voice-based attacks is to use biometric authentication, such as fingerprint scanners, facial recognition or iris scanners.