Cybersecurity Researcher Demonstrates Acoustic Keystroke Logging with 100% Accuracy Using Machine Learning

The video features cybersecurity researcher David Vonin discussing his work on acoustic keystroke logging, a technique that uses machine learning to identify keystrokes based on the sound of keypresses captured via microphones, including during Zoom calls. The method achieved 100% accuracy when trained on a single known keyboard, leveraging subtle acoustic differences caused by wear and tear on frequently used keys (e.g., the letter 'E'). Vonin demonstrated two attack vectors: one targeting a specific keyboard and another using multiple known keyboards, with the latter including a test conducted over Zoom. The research employed PyTorch to train a neural network from scratch, converting audio wave files into spectrogram images to detect unique frequency patterns for each key. Defenses proposed include noise interference (e.g., running water or inverse sound emission) and hardware-based authentication like U2F keys. The discussion also touched on broader AI security risks, including adversarial techniques to disrupt pattern recognition. Vonin emphasized the dual-use nature of AI, highlighting both its potential for malicious exploitation and beneficial applications like drug discovery.