A federated transfer learning approach for surface electromyographic hand gesture recognition with emphasis on privacy preservation

Recently, surface electromyographic (sEMG) hand gesture recognition faces a serious challenge of limited training data in various scenarios. Numerous efforts have been made to address this issue by leveraging data from other subjects. However, the utilization of sensitive sEMG data from other subjects raises the risk of privacy leakage and data security. In this study, a novel federated transfer learning approach is proposed, aiming to use data from other subjects to enhance recognition accuracy while protecting the privacy of these data. At first, a hybrid model incorporating a convolutional neural network and self-attention has been introduced, which has strong gesture recognition ability. Then, a two-stage federated transfer learning framework is proposed, including federated pre-training stage and personalized fine-tuning stage. In the first stage, a federated learning strategy is used to pre-train a knowledge-sharing hybrid model without revealing raw sEMG data from other subjects. In the second stage, a transfer learning strategy is applied to fine-tune the hybrid model according to the characteristics of the target subject. Experimental results show that the proposed method not only addresses the challenge of insufficient data by federating with other subjects while prioritizing privacy preservation, but also helps to enhance the personalized adaptation of shared knowledge to the target subject. Comparative analyses against local training and traditional federated learning reveal significant accuracy improvements of up to 238.52% and 124.01%, respectively, underscoring the efficacy of our proposed method.