Personalized Federated Learning with Dual Collaborative Information for Data Heterogeneity

Personalized federated learning has gained significant attention due to its effectiveness in addressing data heterogeneity and achieving personalization. To train a model that performs well on local datasets, the key is to enable more effective knowledge sharing between clients and servers without compromising data privacy. Most current approaches aggregate partial model on central server for collaborative learning. However, the information conveyed through model parameters is limited. In this work, we propose a dual-information guided personalized federated learning approach called FedDCI. To obtain more collaborative information, we capture information from both model Parameters and data features. By implementing prototypes similarity-based aggregation, we generate client-specific global models and prototypes, effectively mitigating the challenges posed by data heterogeneity in Federated Learning. We conduct extensive experiments in both label skew and feature skew settings using four benchmark datasets. Experimental results demonstrate that FedDCI achieve accuracy improvements over existing methods in complex heterogeneous scenarios.