Federated Learning (FL) is a promising distributed learning method that enables multiple clients to collaboratively train a shared model without sharing their raw data thus preserving privacy. However, in practical implementations, client data are typically non-independent and identically distributed (non-IID). This resulting in heterogeneous learning dynamics and unequal benefits across participants. Improvements in average global performance can mask performance degradation for disadvantaged clients, highlighting a structural fairness challenge in FL. This thesis argues that achieving fairness under non-IID FL requires explicit understanding and modeling of client behavioral heterogeneity rather than uniform aggregation of client updates. 

In addressing the issue of fairness in FL under data heterogeneity, the thesis first studies and analyzes clients' deviating behavior during the federated training process. An eccentricity-based approach is introduced to quantify deviations in local models and data representations within the global model, enabling systematic identification of atypical contribution and benefit patterns. The insights gained lay the foundation for our further research into developing novel, fairness-aware FL solutions for heterogeneous, distributed learning setups.

Then it proposes a fairness-aware aggregation framework called FeDABoost that adapts client influence based on local performance signals. By dynamically weighting client updates and adjusting local optimization to emphasize hard examples, the method reduces disparities across heterogeneous clients while maintaining competitive global performance. Later, the thesis introduces DEFFT, a clients distribution-aware framework that models latent similarities among clients through persistent grouping based on label distributions. Cluster-level models and hierarchical knowledge distillation integrate inter-client structure into the learning process, enhancing fairness metrics along with overall accuracy.

Across multiple benchmark datasets, the proposed approaches demonstrate that a principled way to modeling heterogeneity can lead to measurable improvements in fairness without compromising global performance. The three discussed studies together establish a structured framework for mitigating unequal benefits in FL under non-IID data distributions.