Personalized Federated Learning via Classifier Similarity-based Clustering and Bi-Level Optimization

Federated learning (FL) has emerged as a promising strategy for addressing privacy concerns in today's data silos. However, the performance of traditional FL may suffer from non-independent and identically distributed (Non-IID) conditions. To overcome this challenge, some recent works have explored personalized federated learning (PFL). However, they either consider a single global model for knowledge aggregation, which is a potential bottleneck to handle highly divergent scenarios, or formulate multiple global models, which requires considerable computational overhead. Some of them even have a risk of privacy leakage. In this paper, we propose a novel PFL algorithm named PFedCSCBO, which achieves Personalized Federated Learning via Classifier Similarity-based Clustering and Bi-Level Optimization. It exploits affinity propagation (AP) clustering algorithm based on local classifier similarity to transform the Non-IID condition into multiple IID problems, and employs bi-level optimization for personalization. Furthermore, we introduce an initialization for newcomers to enhance the practicality of PFedCSCBO. Experimental results on several real-world datasets with significant statistical divergence in both convex and non-convex cases show that, our proposed PFedCSCBO outperforms the state-of-the-art methods across various metrics. © 2025 Elsevier B.V.