OpenKeystroke: An Open Research Platform for Acoustic Side-Channel Experiments and Privacy-Defense Evaluation

Claims about acoustic keystroke side channels, attacks that infer typed content from the sound of a keyboard, are difficult to compare because studies differ in recording conditions, data splits, and how defenses are tested. This paper presents OpenKeystroke, an open source research platform for authorized acoustic side-channel experiments and privacy-defense evaluation. The platform measures how much class information survives in short acoustic transients, and how much signal transformations (defenses) reduce it, under a strict evaluation protocol: session, device, and environment-disjoint splits with leakage checks, train-only fitting of preprocessing statistics, validation-only model selection, cluster-bootstrap confidence intervals, and paired comparisons across repeated seeds. Defenses are labelled effective only when they produce a significant accuracy drop, over at least three seeds, against both a static attacker and an adaptive attacker who knows the defense is applied. The platform ships eight classification models and six defenses behind one experiment lifecycle shared by a command line interface, a Python API, and an interactive dashboard. On a synthetic benchmark, headline accuracy degrades materially under domain shift, and several defenses that appear strong against a static attacker largely fail against an adaptive one, which is exactly the gap the platform exists to expose. OpenKeystroke operates only on predefined, non-sensitive acoustic categories in controlled or explicitly consented data. It does not reconstruct typed text, passwords, or messages.