Acoustic Side Channel Attacks: Keystroke Classification from Microphone Recordings

Acoustic side-channel attacks exploit information leakage from physical signals such as sound, timing, or electromagnetic emissions. Keyboards are a relevant target because each press produces a short transient with repeatable spectral structure, enabling classification under suitable recording conditions.

End-to-end flow: record audio → bandpass filter → RMS envelope → peak detection + debounce → onset-aligned window extraction → MFCC(+Δ,+ΔΔ) → scaling → SVM prediction → decoding.

Segmentation uses bandpass filtering (100–8000 Hz) and RMS energy envelope (20 ms frames, 10 ms hop), smoothed by a moving average. Peaks are detected via thresholding and debounced to avoid double-counting. Extracted windows are onset-aligned with a fixed pre/post region for consistency between offline and live.

Each keystroke window is converted to MFCC (20 coefficients) with Δ and ΔΔ, and summarized via per-coefficient mean and standard deviation. A StandardScaler is fit on training data only, and an RBF SVM performs multi-class classification. Session/group-aware splitting controls leakage.

Word decoding combines: per-position probabilities (softmax over SVM margins), beam search over letter sequences, and dictionary snapping with confusion-aware substitution costs derived from observed misclassifications. Sentence mode segments by predicted SPACE and decodes each word chunk with relaxed constraints.

A Streamlit GUI provides three demo modes (letter vote, fixed-length word, sentence) with per-mode tuning controls for decoder temperature, beam width, and top-K expansion. Sentence mode adds sensitivity (threshold K) and debounce controls.

The player below attempts to load the YouTube embed. If your environment blocks embedded playback, use the button to open the video directly.