FedHMIR: Unified Framework for Federated Human-Machine Synergy in Personalization-Generalization Balancing Identity Recognition

As device-free identity recognition (IR) gains popularity and the demand for the Internet of Things (IoT) continues to grow, a new-era IR system featuring multiple distributed recognition devices and edge servers faces two main challenges: model adaptability and balancing the personalization of devices with the generalization of the system. This research introduces FedHMIR, a federated framework designed to simultaneously address these challenges by harmonizing human-machine collaboration with personalization-generalization trade-offs. The proposed framework features a human-machine cooperative online internal update mechanism, leveraging reinforcement learning to maintain the adaptability of personalized local IR models. To counter overfitting and enhance the generalization of the overall IR system, an external update process incorporating a confidence index is introduced. Additionally, the framework employs asynchronous internal and external update procedures to effectively balance personalization and generalization between local and global models. Finally, extensive experiments on three diverse real-world datasets demonstrate the effectiveness and advantages of FedHMIR compared to state-of-the-art baselines.