Personalized Cluster-Based Model Adaptation for Non-IID Federated Systems

The emergence of federated learning (FL) addresses the critical challenge of “data silos” caused by privacy regulations, policy constraints, and commercial confidentiality that restrict cross-source data sharing. While FL enables collaborative model training without raw data exchange, studies demonstrate that global model performance heavily relies on consistent participation of high-quality clients. However, real-world scenarios often involve heterogeneous data quality across clients, leading to unfair treatment during training and potential client dropout. This creates a participation paradox: despite FL's privacy preservation, the substantial computational costs discourage clients lacking fair compensation from sustained engagement. Motivating active and equitable collaboration therefore remains a pivotal challenge. Current fairness-enhancing approaches exhibit notable limitations. Shapley value-based contribution measurement methods, while effective for horizontal FL, fail to generalize across federated architectures. Reputation-based iterative frameworks attempt to quantify client value but suffer from subjective evaluation metrics and lack rigorous data quality assessment protocols. To overcome these limitations, we propose FedPC - a dual-phase fairness optimization framework. Server-side label-aware clustering aggregates clients with semantic similarity, while client-side personalized model training ensures task-specific adaptation. Evaluations on EMNIST and CIFAR-10 demonstrate FedPC's superiority: it achieves higher fairness scores than baseline FL methods (e.g., FedAvg, FedProx) while maintaining competitive accuracy, striking an optimal balance between collaborative equity and model efficacy.