SKAID: A Realistic Dataset for Acoustic Side-Channel Attacks with Synchronized Keyboard Audio and Keystroke Logs for AI-based Inference

Acoustic side-channel attacks (ASCAs) exploit the subtle acoustic signatures produced by keystrokes to reconstruct typed content. Machine learning models have achieved remarkable inference accuracy in controlled environments, raising pressing concerns about user privacy and system security. Yet, progress in acoustic side-channel attack research has been limited by the availability of high-quality, realistic datasets that combine synchronized audio and key event logs with contextual demographic data.
To address this gap, we introduce a new methodology for synchronized keystroke and acoustic data collection to capture keystroke-level logs, raw keyboard audio, transcribed text, and user demographics in naturalistic conditions. We demonstrate the feasibility and efficiency of the proposed methodology by building SKAID (Synchronized Keystroke and Acoustic Inference Dataset), a high-quality dataset in a real-world setting. Unlike prior resources restricted to isolated characters or laboratory typing, SKAID includes both structured email-style text and unstructured free typing tasks, captures long-form keystroke event logs at millisecond resolution, and keyboard audio and demographic metadata. Beyond advancing ASCA research, SKAID raises important ethical considerations regarding dual-use risks; therefore, we propose guidelines for the responsible dissemination of such research. By releasing SKAID, we contribute not only a novel, high-fidelity dataset but also a reproducible methodology that strengthens AI-driven cybersecurity research and informs the design of countermeasures against acoustic emanation attacks.