New acoustic attack steals data from keystrokes with 95% accuracy

Researchers trained a deep learning model to steal data from keyboard keystrokes using audio recordings with up to 95% accuracy. They recorded keyboard presses and generated spectrograms to train an image classifier to identify keys. Testing involved recording sounds from a phone near the keyboard or through Zoom. The model achieved 95% accuracy from phone recordings and 93% from Zoom. While some mitigations were suggested, like altering typing styles, the attack was still effective even against silent keyboards. This raises serious concerns about data security as sensitive information could be leaked through acoustic side-channel attacks.