Protecting against data theft has become an increasingly challenging task in today's digital world. Techniques such as phishing are increasingly sophisticated to obtain user personal information. Cybersecurity companies alert this week of a new development that puts device privacy in jeopardy: an artificial intelligence (AI) capable of stealing passwords simply by listening to the sound of the keys while typing.
Joshua Harrison, Ehsan Toreni and Maryam Mehrnezhad are the Cornell University researchers who have developed this innovative AI. Their study details a method that involves pre-training artificial intelligence to recognize a user's typing patterns. The result has been all the more surprising, since they have managed to predict with 95% accuracy what a user is typing on their computer.
The AI can pick up on the distinctive sounds made by the keys on a keyboard. Of course, it should be noted that the training carried out is not enough to cover all the types of keyboards that exist and, for its optimal functioning, it is necessary to train for each model. Then you have to link the recognition results with the corresponding keys.
This procedure can be done in several ways. As the researchers indicate in their study, it can be done through a microphone or remotely using apps like Zoom. In the latter case, the fact that the accuracy of the AI decreased slightly, up to 93%, stands out.
The team also tested this AI using a MacBook Pro. To train the system, they pressed 36 keys (letters and numbers) 25 times each. Thus, the AI model was able to associate the individual sounds with the corresponding characters. Mechanical keyboards performed better than membrane keyboards, as they produce more audible and distinctive sounds.
Although the sound waves generated by these sounds are subtle, they were distinctive enough to be associated with specific keys. The result of this examination allowed the researchers to achieve highly accurate text input recognition.
On the other hand, the experts also detected that there are some techniques that can drastically decrease the recognition accuracy. If the typing style is changed or software that generates background sounds is used, the AI has difficulty capturing the audio of the keys. In this case, the system's accuracy dropped to between 64% and 40% when using touch typing.
With these results in hand, the researchers advise exploring methods that alter the computer's audio input rather than opting for quieter keyboards.
