FedEAS cuts synthetic data generation budget by 94.1% in federated learning
Near-perfect accuracy with a fraction of the generation cost—how FedEAS does it.
Label skew in federated learning causes client drift and degrades global model accuracy. While synthetic data augmentation can rebalance skewed distributions, full class balancing is computationally expensive. Researchers from the field present FedEAS, a budget-aware policy that determines both how much synthetic data each client generates and where those samples are allocated. Instead of a fixed global budget, FedEAS computes an entropy-adaptive per-class generation budget from each client's local label distribution. This joint decision dramatically reduces waste while preserving accuracy gains.
The experimental results on CIFAR-10 and CIFAR-100 show that FedEAS recovers most of the accuracy improvement of exhaustive class balancing while cutting the total generation budget by 94.1%. Compared to a uniform budget allocation across clients, FedEAS improves global accuracy by up to 18.82%. This work is especially relevant for privacy-sensitive and resource-constrained federated learning deployments, where communication and computation are limited. The paper is available as arXiv:2607.06616 and preprinted on July 7, 2026.
- FedEAS reduces synthetic data generation budget by 94.1% compared to full class balancing.
- Outperforms Uniform allocation by up to 18.82% on CIFAR-10 and CIFAR-100.
- Uses entropy-adaptive per-class budgets to decide how much and where to generate synthetic samples.
Why It Matters
Enables efficient, accurate federated learning in resource-limited settings by slashing synthetic data costs.