WHERE to Generate Matters: Budget-Aware Synthetic Augmentation for Label Skewed Federated Learning

FedEAS introduces a policy for budget-aware synthetic augmentation in federated learning to combat label skew and client drift. By using entropy-adaptive per-class generation budgets based on local distributions, it recovers accuracy gains while reducing computation costs. Separately, new research benchmarks fairness-aware learning on differentially private synthetic tabular data. This study evaluates how Differential Privacy affects fairness interventions, addressing the conflict between privacy-preserving data analysis and the mitigation of discrimination against underrepresented demographic groups.