AI researchers claim 93% accuracy in detecting keystrokes over Zoom audio

Researchers have developed a deep learning model that interprets remote keystrokes with high accuracy based on the sound profiles of individual keys. This technique poses security concerns, especially in quieter public areas where laptop keyboards might be recorded.

Deep Learning and Remote Keystrokes:

Three researchers trained a deep learning model to interpret remote keystrokes by recording their sounds.

They achieved over 90% accuracy by identifying the unique acoustic profiles of individual keys.

Details from the Research Paper:

The paper, titled "A Practical Deep Learning-Based Acoustic Side Channel Attack on Keyboards," was authored by UK researchers Joshua Harrison, Ehsan Toreini, and Marhyam Mehrnezhad.

The researchers argue that the combination of machine learning, microphones, and video calls puts keyboards at an increased risk of being recorded, especially in quiet public places.

Laptops, particularly, are vulnerable due to their non-modular, uniform keyboards with similar sound profiles across models.

Previous Research on Keylogging:

Past attempts to keylog VoIP calls reached up to 91.7% top-5 accuracy on Skype in 2017 and 74.3% accuracy in 2018.

Using a "hidden Markov model" (HMM) to predict likely next letters improved the accuracy of a previous side channel study from 72% to 95% when applied to dot-matrix printers.

Technological Advances:

The Cornell team believes they are the first to harness recent advancements in neural network technology, such as self-attention layers, for this type of audio side channel attack.