Personalized federated learning with global information fusion and local knowledge inheritance collaboration

Traditional federated learning has shown mediocre performance on heterogeneous data, thus sparking increasing interest in personalized federated learning. Unlike traditional federated learning, which trains a single global consensual model, personalized federated learning allows for the provision of distinct models to different clients. However, existing federated learning algorithms solely optimize either unidirectionally at the server or client side, leading to a dilemma: "Should we prioritize the learned model's generic performance or its personalized performance?" In this paper, we demonstrate the feasibility of simultaneously addressing both aspects. Concretely, we propose a novel dual-duty framework. On the client side, personalized models are utilized to retain local knowledge and integrate global information, minimizing risks associated with each client's experience. On the server side, virtual sample generation approximates second-order gradients, embedding local class structures into the global model to enhance its generalization capability. Utilizing a dual optimization framework termed FedCo, we achieve parallelism of global universality and personalized performance. Finally, theoretical analysis and extensive experiments validate that FedCo surpasses previous solutions, achieving state-of-the-art performance for both general and personalized performance in a variety of heterogeneous data scenarios.