Budget-Aware Synthetic Data Augmentation Improves Federated Learning Accuracy.

Federated learning often suffers from "client drift" and reduced global accuracy due to imbalanced data distributions across different clients, known as label skew. While generating synthetic data can help balance these distributions, fully balancing all classes can be computationally expensive. Researchers have developed FedEAS, a new strategy that intelligently allocates budgets for synthetic data generation. Instead of a fixed total budget, FedEAS calculates an entropy-adaptive budget for each client and each class, determining both the quantity and destination of generated samples. This approach allows the total generation budget to emerge dynamically. The method demonstrates substantial efficiency improvements, achieving most of the accuracy benefits of full class balancing with a 94.1% reduction in generation budget.