A research team from Cornell University has created a system that allows training an AI model for find out what is written on a computer from only the sound that the keys make when pressed to write. The team that did it, made up of researchers Joshua Harrison, Ehsan Toreini and Maryan Mehrnezhad, have already published an article in which they offer details of the work they have done, which cybercriminals could use to launch an attack and steal user credentials. access to teams and accounts. Also to obtain sensitive information with which the objective of the same is working.
By recording the sound of the keystrokes to train the model, the researchers have managed that, once the process is finished, the AI model deduces what has been typed with a computer with 95% accuracy. Of course, when Zoom has been used to train the system, the accuracy of the model in handwriting recognition dropped by 2%, to 93%.
Apart from this, it must be taken into account that the system does not work with any keyboard, but must be trained with a specific keyboard, with references to which letter each key corresponds to and its sound. This can be done locally, with the help of a microphone, or remotely. In the latter case, to record the sound of keystrokes, an application such as Zoom is used.
The team of researchers has demonstrated the project using a MacBook Pro computer, in which they pressed 36 keys 25 times each. This was taken as a basis for the Artificial Intelligence model to recognize which character is associated with a particular key sound. It may seem difficult to distinguish them in this way, but there are some differences in the sound waves that are produced when pressing each one of them that, although they are very subtle, are enough to recognize each letter with a fairly high level of precision.
It is therefore a system that could be used to launch cyberattacks and that allows credentials to be stolen and guess what the target of an attack writes. Of course, the system has many weak points. Aside from the need to use a particular keyboard for training the AI model, steps can be taken to reduce the accuracy of the system. Among them is the change in the writing style of whoever uses the keyboard.
Touch typing also reduces recognition accuracy by 40% to 64%. And if environmental noise is produced by software capable of producing white noise or the sound of other keys, with the noise made when typing on the keyboard, the accuracy of the model is reduced to a minimum.
With everything, the attack that could be launched with this system would work well with mechanical keyboards They have a fairly powerful and perceptible sound. But the model is not only effective with this type of keyboard. It is also with membrane keyboards, since they also produce the sound necessary for the training of the Artificial Intelligence model. Therefore, the best way to avoid the attack is not to change your keyboard and get a silent one, but to implement a software-based solution on your computer.
