Domain generalized federated learning for Person Re-identification

In the field of Person Re-identification (ReID), addressing the demands of practical applications in diverse and uncontrollable unseen domains necessitates a focus on Domain Generalization (DG). However, when tackling DG for human-related tasks, the growing awareness of privacy introduces new challenges. Privacy concerns often prevent the sharing of local datasets for global learning, and this limitation in data sharing can impair the generalization ability. Therefore, it becomes imperative to address domain generalization under the constraint of privacy protection. This paper delves into a novel and challenging domain generalization problem that incorporates privacy concerns. We propose a new generalizable ReID network that integrates decentralized learning from non-shared private training data. To mitigate domain variations among clients, we introduce a dynamic aggregation strategy for learning a domain-invariant server model. This strategy adaptively weights clients, guided by domain-invariance principles. To ensure generalization ability with limited client data, we present a domain compensation network. This network augments fictitious domains in the model design to simulate unseen testing situations. In the process of generating fictitious domains, we integrate diversity to avoid meaningless generations, and we constrain fidelity to preserve discrimination. Extensive experiments demonstrate the effectiveness of our method in enhancing generalization ability and privacy protection. Our approach achieves competitive performance on multiple widely used benchmarks.