Researchers found a fresh way of hacking that makes use of artificial intelligence (AI) by analyzing people’s keyboard input sounds. This is accomplished by using an AI that has been trained to recognize the acoustic profile of keyboard keys, which includes features such as sound frequency, amplitude, and timing for each key press.
A research paper by Joshua Harrison et al. presents a practical implementation of state-of-the-art deep learning model. Whereas it is able to “classify laptop keystrokes, using a smartphone integrated microphone.”
When trained on keyboard data taken by a nearby smartphone, the classification model obtained an astonishing 95% accuracy, the best precision achievable without the use of a language model.
In an experimental situation, researchers assessed the model’s performance during a Zoom call by recording keyboard sounds with the laptop’s built-in microphone. The AI model faithfully reproduced the recorded keystrokes with a remarkable 93% accuracy. Similarly, in a comparable testing using Skype, the model achieved an accuracy level of around 92%.
Recent developments in the proliferation of microphones located within the acoustic range of keyboards have made this novel hacking tactic more exploitable.
To validate their hypothesis, the research team systematically pressed individual keys on a MacBook Pro 25 times each, aiding the system in discerning discernible patterns. Subsequently, an iPhone 13 mini, positioned 17 cm away from the keyboard, was employed to capture keystroke audio for the initial test.
For further analysis, the researchers recorded laptop keystrokes over Zoom using the internal microphones of the MacBook.
It is cautioned by the researchers that this emerging technique, leveraging the synergy of AI, microphones, and video conferencing, engenders a heightened threat to keyboard security.
Nonetheless, it is emphasized that the AI model necessitates distinct training for each keyboard type to effectively associate each keystroke with its corresponding character.
To mitigate the susceptibility to such attacks, the study advocates for altering one’s typing mannerisms and recommends the use of randomized passwords incorporating diverse cases. Moreover, the adoption of biometric authentication methods like fingerprint or facial recognition is encouraged as an alternative to conventional typed passwords.