FedEAS Optimizes Synthetic Data Generation for Federated Learning

Federated learning (FL) often suffers from performance degradation due to label skew, where data labels are unevenly distributed across client devices. While synthetic data augmentation can help balance these distributions, fully balancing all classes is computationally expensive. This research proposes FedEAS, a novel policy designed for budget-aware synthetic augmentation that intelligently allocates generation resources. FedEAS assigns each client an entropy-adaptive, per-class generation budget, which is calculated based on its local label distribution. This budget determines both the quantity of synthetic data each client generates and the specific classes for which it generates samples. This dynamic allocation strategy allows the total generation budget to emerge from client-specific needs rather than being fixed upfront. Experiments on CIFAR-10 and CIFAR-100 datasets demonstrate that FedEAS recovers most of the accuracy benefits of full class balancing while reducing the overall generation budget by an impressive 94.1%. It also outperforms uniform allocation strategies by up to 18.82% for the same total budget.