SWiST DeepNet: Spectral-Temporal Deep Network for Detecting Resonance-Based Attacks on MEMS Inertial Sensors

Microelectromechanical systems (MEMSs)-based inertial sensors, such as accelerometers and gyroscopes, are fundamental to the operation of cyber-physical systems (CPSs), enabling real-time motion tracking and orientation estimation. However, their susceptibility to resonance-based acoustic interference can lead to false measurements, corrupting sensor outputs and potentially destabilizing the systems that rely on them. To detect such threats, this article proposes a sliding-window-based deep learning framework that uses temporal- and spectral-domain features to identify resonance-induced signal injections. The model is evaluated under diverse and realistic operating conditions, including both static and dynamic motion, and across three defined attack levels: soft, medium, and hard. The experimental analysis compares the performance of temporal, spectral, and a fusion of spectral- and temporal-domain features for attack detection across different window lengths. The proposed sliding window spectral-temporal deep network (SWiST DeepNet) architecture reliably detect attacks even at low intensities down to a 10-dB attack-to-signal ratio (ASR), significantly enhancing the resilience and trustworthiness of CPS against resonance-based attacks. Our datasets are available at https://github.com/cmsahu97/RAiS-dataset.