Cluster Federated Learning with Intra-Cluster Correction

Federated learning has emerged as an essential technique of protecting privacy since it allows clients to train models locally without explicitly exchanging sensitive data. Extensive research has been conducted on the issue of data heterogeneity in federated learning, but effective model training with severely imbalanced label distributions remains an unexplored area. This paper presents a novel Cluster Federated Learning Algorithm with Intra-cluster Correction (CFIC). First, CFIC selects samples from each cluster during each round of sampling, ensuring that no single category of data dominates the model training. Second, in addition to updating local models, CFIC adjusts its own parameters based on information shared by other clusters, allowing the final cluster models to better reflect the true nature of the entire dataset. Third, CFIC refines the cluster models into a global model, ensuring that even when label distributions are extremely imbalanced, the negative effects are significantly mitigated, thereby improving the global model’s performance. We conducted extensive experiments on seven datasets and six benchmark algorithms. The results show that the CFIC algorithm has a higher generalization ability than the benchmark algorithms. CFIC maintains high accuracy and rapid convergence rates even in a variety of non-independent identically distributed label skew distribution settings. The findings indicate that the proposed algorithm has the potential to become a trustworthy and practical solution for privacy preservation, which might be applied to fields such as medical image analysis, autonomous driving technologies, and intelligent educational platforms.