FedGPA: a multi-level aggregation framework for federated learning with collaborative optimization of generalization and personalization

Abstract

Federated Learning (FL) is an innovative approach that balances data privacy protection and distributed model training, effectively addressing the challenges posed by data silos and the growing demand for privacy preservation. However, the heterogeneity of client data, including Non-Independent and Identically Distributed (Non-IID) data and imbalanced datasets, poses significant challenges in balancing global generalization with local personalization. Existing approaches to optimizing aggregation strategies for personalized federated learning remain insufficient, as they fail to fully account for the impact of data heterogeneity and hierarchical model features, thereby limiting personalization and model accuracy improvements. This paper introduces a novel framework, FedGPA, which integrates the global model, clustered models, individual optimal models, and individual local models through an aggregation mechanism leveraging the hierarchical structure of deep neural networks. The framework incorporates an adaptive weight matrix that dynamically adjusts feature contribution ratios across different layers. Furthermore, a hyperparameter is introduced to control the aggregation scope, enabling lower-layer features to emphasize generalization while higher-layer features focus on task-specific personalization. Component-wise ablation studies are further conducted to validate the contribution of each fusion branch. Experimental evaluations on benchmark datasets including MNIST, SVHN, and CIFAR-10, demonstrate that FedGPA outperforms existing methods.