You might want to be careful the next time you use your computer while on a call because AI systems are getting smarter and more sinister every passing day.
AI System Trained To Decode Keystrokes On Laptops By Listening To What Users Type
A group of researchers from the U.K’s Cornell University have trained an artificial intelligence algorithm to successfully decode keystrokes by overhearing them through video or conference calls. This means that the AI can find out what you’re typing on your computer even if the screen or the keyboard is not visible.
The scientists behind the study – Joshua Harrison, Ehsan Toreini, and Maryam Mehrnezhad – recorded the sound of each key using a smartphone-integrated microphone and made the deep learning model listen to it.
The team trained the model to recognize the unique patterns with which users press the keys on a Macbook keyboard, including the sound, intensity, and time it took to press and release each key.
Harrison and his colleagues put the Macbook Pro and the iPhone 13 at a distance of 6.5 inches and then pressed each of the 36 individual keys 25 times, while it was all being recorded on the smartphone Then the scientists started a video conference call on Zoom and Skype and recorded the keystrokes using the Macbook’s built-in microphones.
The Algorithm Has An Accuracy Rate Of Over 90% Without Using Any Language Model
According to the researchers, when trained by listening to the keystrokes recorded by a smartphone, the AI could recognize the keys with 95% accuracy, which was the highest success rate observed without the use of a language model.
When trained by listening to keystrokes recorded on the laptop while using Zoom, it was able to reproduce the exact keys with 93% accuracy. The model was found to be 92% accurate in identifying the keystrokes when tested on Skype.
Even when the AI was not able to determine the character that was being typed, it was off by only one key, making it easy for potential bad actors to go around the error. The researchers claim that cybercriminals can make good use of the new method due to the fact that the number of microphones within an acoustic range of keyboards has increased vastly with technological advancements in the past decade.
Although it is with a good intention of improving audio quality, the cybersecurity threat integrated microphones pose cannot be ignored.
Scientists warn that this new technique has presented the greatest threat to using physical keyboards.
However, the AI system is not able to work the same way with every keyboard because it has to be trained separately to understand the keystrokes of different computers. Also, the model has to be presented with additional references to understand the features of the key that is being pressed and what character it corresponds to.
Researchers List Out Ways In Which Users Could Mitigate The Threat Of Hacker AI.s
But, it is not the end of the world for computers. The study found that users can mitigate the threat of these kinds of attacks by just changing their typing style when typing sensitive information.
According to the researchers, touch typing reduced the AI’s keystroke recognition accuracy by 60% to 40%. Playing a software-based keystroke audio filter in the background also helped disguise keystroke sounds.
Biometric passcodes like fingerprint or face ID are recommended. The scientists suggest using randomized passwords that feature multiple cases and numerals since large language models like ChatGPT are able to predict succeeding characters to complete words. That means, passwords containing full words are a thing of the past and it’s time to change yours.