Towards Robust, Personalized, and Fairness-Aware Federated Learning

Abstract

Federated Learning (FL) enables collaborative model training without sharing raw data, but traditional aggregation methods such as FedAvg overlook data distribution, quality, and fairness. This limitation becomes critical in real-world settings, where client data is highly non-independent and identically distributed (non-i.i.d.), noisy, or weakly labeled. Simply weighting updates by dataset size can amplify low-quality data, harm generalization, and introduce fairness risks, especially in weakly supervised scenarios like Partial Label Learning (PLL), where label ambiguity can be exploited adversarially. This thesis investigates personalized and fairness-aware FL under increasingly realistic data assumptions. First, we develop pFedMo, a personalized FL algorithm that mitigates data heterogeneity by combining contrastive learning with personalized FL. It reduces local bias through a contrastive aggregation score aligned with a central representation model trained on i.i.d.~data, and enhances convergence through personalized momentum. Next, to account for weak supervision, we develop pFedPLL, which addresses label ambiguity and correlation interference in non-i.i.d.~settings. It preserves local label structure via Label Correlation Isolation and improves prediction accuracy through bi-directional calibration. Finally, to ensure fairness and robustness in weakly supervised FL, we develop FairFedPAPL, a defense framework for federated partial attribute and partial label learning. It detects and mitigates fairness-related attacks by reconstructing representative client data through gradient inversion, allowing effective fairness preservation without compromising privacy. In summary, this thesis presents three algorithms to tackle key challenges in FL, including data heterogeneity, weak supervision, and fairness under non-i.i.d.~settings. Extensive theoretical analyses and experiments confirm their effectiveness and practicality in realistic FL environments.