Behavioral Dataset for Ultrasonic Command Injection Detection in Voice-Controlled Systems

This dataset provides real-world behavioral interaction metadata collected from voice-controlled systems deployed in residential smart-home environments over a 90-day period. It contains cloud-accessible user and device command logs capturing temporal execution patterns, semantic intent information, audio availability indicators, and device context attributes. Anomalous instances are behaviorally labeled based on manifestations consistent with ultrasonic command injection threat models established in prior literature, without performing any physical ultrasonic signal injection. The dataset is intended to support research on voice assistant security, behavioral anomaly detection, and machine learning\u2013based defense mechanisms. No raw audio or personally identifiable information is included, ensuring privacy-preserving data sharing.