Robust and Scalable Federated Learning Framework for Client Data Heterogeneity Based on Optimal Clustering

Abstract

Federated learning is a promising paradigm for applications across a variety of domains. However, there are some challenges that must be addressed in real-world scenarios, particularly the data heterogeneity among participating clients. Most existing studies primarily focus on the issue of non-independent and identically distributed data, but they do not consider the critical aspect of data quality heterogeneity. When low-quality data is contributed by some clients, the efficacy of models trained through the traditional approaches will be significantly compromised. Therefore, we propose ROSCFL, a robust and scalable federated learning framework for client data heterogeneity based on optimal clustering. We first develop a cluster contribution evaluation strategy based on the optimal clustering to quantify the contribution of each cluster. Next, we design a robust model aggregation strategy, which effectively mitigates the impact of low-quality data on the global model by optimizing weight allocation and client sampling. Finally, we introduce a client incorporation mechanism to enhance the scalability of ROSCFL. Extensive experiments have been conducted, and the results demonstrate that ROSCFL achieves strong robustness and scalability, particularly in scenarios wherein data distribution and quality heterogeneity coexist.