2023-08-07 20:43:49
Can you imagine if someone might discover the content of your typing just from the sound of the keys being pressed? UK researchers have developed a language model with up to 95% accuracy which manages to do just that, associating — in a complex way — the sound emitted by typing specific keys, in order to identify what was typed only by the noise.
As echoed by Bleeping Computer, Joshua Harrison, Ehsan Toreini e Maryam Mehrnezhad published the results of research on the topic in an article. They managed to create the model using a MacBook Pro and the microphone of an iPhone 13 mini, positioned 17 centimeters from the laptop.
The sound emitted when pressing each of the 36 keys on the MacBook’s keyboard was recorded 25 times each, producing different sound waves and spectrograms, as we can see below. With this, it was possible to notice the differences in the sound generated by each key, in order to be able to identify them individually using the sound and, thus, train the algorithm.
The spectrogram images were used to train an image classifier called CoAtNet. The process involved some experimentation time, learning rate and image separation parameters so that high precision was achieved in the algorithm created by the group of researchers.
When recording audio directly from the environment where the MacBook was, the accuracy was 95%, while with the sound of a phone call Zoomthe number dropped a little, but still remained at the high level of 93%while in connections on the Skype the proportion of correct answers for the language model was 91,7%.
Thus, being in a group video call, for example, someone who wanted to find out what you are typing would need to record your typing and associate the sound with the written content. It would be necessary to calibrate the model to each person’s keyboard to then carry out the described procedure, exploring the differences between the sounds of each key. Even on silent keyboards, creation works, according to the researchers.
It is worth remembering that the risk of this type of algorithm does not only exist in relation to people nearby or who are on calls, but also, for example, with its combined use with malware. Someone might then hack into your iPhone’s or even Mac’s microphone and then capture your typing sound in search of passwords.
In the article, the researchers recommended, as a way of mitigating the problem, varying the way used to type, especially passwords and/or sensitive information, or using stronger passwords. It is also possible to use software with audio filters, which reproduce white noise or another type of sound to mask typing.
The other most obvious tip is, when you’re on a call, don’t type passwords or mute your microphone when you do. Saving them in a password manager and using autofill features is also something that counteracts the use of language models like the one developed by the researchers. We can’t forget, yet, the Password Keys (Passkeys), which completely eliminates the need to type passwords.
Who knew that even typing might be a risk, eh? ????
1691456540
#Language #model #discovers #passwords #sound #typing