FedAPI: Privacy-preserving Multi-end Adaptive Personal Identification via Federated Learning

Activity-based personal identification approaches are explored due to fidelity, resilience, and convenience. In an Intelligent Internet of Things (AIoT) environment with multiple identification ends, adaptive personal identification methods with centralized learning are not sufficient to support the imbalanced and non-IID distribution inherent in locally collected data. To learn a more accurate identification model with privacy-preservation by using the decentralized data distributed in different clients, we propose FedAPI, a privacy-protected multi-end adaptive personal identification framework with federated learning. Considering the universality and adaptability of the system, we devise a global update and a local update mechanism. Moreover, the reliability of each identification end is calculated according to the human expert feedback. Experiments are conducted on two activity-based personal identification datasets. Compared to centralized personal identification methods, FedAPI achieves a gain of more than 9% precision.