A Resource-Efficient Cross-Domain Sensing Method for Device-Free Gesture Recognition With Federated Transfer Learning

Emerging Wi-Fi sensing technologies and applications make indoor ubiquitous sensing possible. Deep learning empowered Wi-Fi sensing model enables detection, recognition automatically, without being explicitly programmed. Due to indoor sensing privacy and ubiquitous sensing ability concerns, it is absolutely necessary to conduct in-depth research on Wi-Fi sensing security training and cross-domain sensing issues. To address them, in this paper, we propose to learn domain independent features, distributed model training and localized inference based on federated transfer learning. Moreover, several efficient methods are proposed to provide a distributed edge Wi-Fi sensing scheme with sensing data privacy protection, costing less time, communication, computing and energy resources. We implement the proposed framework and conduct experiments with Widar3.0 datasets to evaluate its performance. The results demonstrate that our framework performs better for cross-domain Wi-Fi sensing while preserving user data privacy and saving resources.