Enhancing fairness in federated learning using client clustering and personalized models

Federated learning (FL) enables collaborative model training across distributed clients while preserving data privacy, but non-independent and identically distributed (non-IID) client data can amplify algorithmic bias and raise significant fairness concerns. This paper presents a systematic empirical evaluation of clustered federated learning (CFL) strategies, analyzing accuracy and fairness across different numbers of clusters (k) and controlled levels of heterogeneity generated via Dirichlet partitioning with concentration parameter \alpha. By varying these parameters, we study behavior from extreme class imbalance to near-IID client distributions under different demographic contexts. We compare four representative strategies: FedAvg, IFCA, FL+HC, and WeCFL, on the COMPAS (race-based fairness) dataset and validate key patterns on the Adult (gender-based fairness) dataset, using predictive accuracy and five complementary fairness metrics. Overall, clustering can improve both fairness and accuracy in highly heterogeneous regimes, but benefits diminish as data become near-IID and some methods (notably FL+HC) exhibit instability under specific configurations; one-way ANOVA and Tukey HSD confirm that strategy choice significantly affects accuracy, while increasing clusters beyond k=5 yields limited additional gains, providing practical guidance for deploying heterogeneity-aware FL systems that seek measurable fairness improvements without sacrificing performance.