Fairness in federated learning through client clustering and model personalisation

Abstract

Federated learning (FL) is a pivotal training paradigm of machine learning models on distributed clients to put data privacy into practice. Nonetheless, those inherent non-independent and identically distributed (non-IID) data patterns of different federated clients can seriously magnify algorithmic bias as well as fairness issues, especially among undersized demographic fractions. This dissertation examines how fairness measures are influenced by clustered federated learning (CFL) approaches through an in-depth empirical analysis of several datasets and demographic settings.

The results show inherent conflicts among contrasting fairness goals: accuracy parity- optimizing methods can simultaneously degrade predictive parity, and individual unfairness- decreasing methods can elevate group-level disparities. These compromises become more pronounced as data heterogeneity increases, with α (the lowest concentration) equal to 0.01, resulting in disparate impact ratios as high as 0.53 for FL+HC, which reflects systematic protection- class discrimination. Cross-dataset validation validates that fairness patterns generalize across demographic settings, such that the race-based fairness of COMPAS shows similar trade-offs to the gender-based fairness of the ADULT dataset. Yet effect magnitudes and directions are different, indicating context-contingent deployment approaches are necessary. Optimization of hyperparameters achieves 2-4% accuracy gains and 8-12% fairness improvements, proving that optimal tuning is necessary to meet fairness goals. This work presents important new insights to effectively implement balanced federated learning systems in high-stakes situations like employment screening, credit scoring, and criminal risk assessment. We develop evidence-based recommendations to practitioners on approach choosing, cluster setting, and data allocation requirements that must be used to improve fairness measurably without compromising model performance.

This work additionally delivers an extensive empirical study of fairness in Clustered Federated Learning (CFL), analyzing how various clustering strategies and cluster setups affect both accuracy and fairness in the presence of heterogeneous client data. Four federated learning methods, FedAvg, IFCA, WeCCFL, and FL+HC, were evaluated under different levels of data heterogeneity and for multiple numbers of clusters (k = 3, 5, 7). Statistical significance analyses, including one-way ANOVA and Tukey HSD post-hoc tests, verified that the observed variations in accuracy and fairness across methods and cluster sizes were not due to random noise but instead represented real algorithmic differences. In particular, IFCA and a configuration with k = 5 clusters consistently achieved the strongest results, yielding higher accuracy and better fairness across several evaluation criteria. These outcomes demonstrate that, under non-IID conditions, fairness and model performance can be simultaneously improved when clustering is applied carefully. Taken together, this work contributes methodological guidance, clarifies trade-offs between different fairness notions, and provides actionable recommendations for implementing fair and robust CFL systems in practical deployments.

Date of Thesis submission	2025
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Kenan Matawie (Supervisor) & Bahman Javadi Jahantigh (Supervisor)