This research demonstrates how Large Language Models can significantly improve acoustic side-channel attacks by correcting errors in keyboard sound recordings, making these attacks viable even in noisy real-world conditions.

• Combines spectrograms of keyboard sounds with LLMs to identify keystrokes even with background noise
• Achieves up to 88% accuracy in keystroke identification under noisy conditions
• Leverages LLM capabilities to correct typing errors similar to predictive text
• Creates more viable attack vectors through ubiquitous microphones in everyday devices

These findings highlight critical security vulnerabilities in environments previously considered safe from acoustic eavesdropping, requiring organizations to reassess physical security protocols for sensitive information.

Read the paper: Making Acoustic Side-Channel Attacks on Noisy Keyboards Viable with LLM-Assisted Spectrograms' "Typo" Correction