DPMAS-Net: A Privacy-Preserving Deep Learning Model for EMG-Based Hand Gesture Recognition with Time-Frequency Domain Features

Hand gesture recognition via electromyogram signals plays a crucial role in enhancing human-machine interfaces, with applications in assistive technology, prosthetics, and smart devices. This paper introduces DPMAS-Net, a deep learning model that classifies EMG-based hand gestures while prioritizing patient data privacy using differential privacy techniques. This model effectively maintains a high gesture recognition accuracy comparable to non-private models, tackling the usual tradeoff between privacy and performance. The model has been validated on multiple datasets. This study reveals that, even with privacy-preserving measures in place, the DPMAS-Net's accuracy approaches current state-of-the-art results with privacy and surpasses them with non-privacy (Ninapro DB4 86.4% with privacy and 92.0% without privacy, Biopatrec DB2 89.3%, and the Mendeley Dataset 90.4%). Therefore, by combining deep learning techniques with differential privacy, the DPMAS-Net model presents a significant advancement in EMG signal classification for hand gesture recognition.