Heterogeneous FL via Active-Passive Collaboration

Federated Learning (FL) is a privacy-preserving machine learning framework that often faces performance challenges due to data and system heterogeneity. Current client selection strategies improve heterogeneous FL performance but face two main challenges. Our preliminary experiments show that data heterogeneity arises not only globally but also within individual clients due to local data volume differences. However, typical server-side client selection strategies cannot fundamentally address accuracy declines from local data heterogeneity. Additionally, clients with high-quality data typically require longer training times, creating a trade-off between accuracy and training duration, and many strategies exclude these clients due to time delays, further reducing accuracy. To address these issues, we propose FCH-FL, enhancing heterogeneous FL through active-passive collaboration. Our method includes a client-side data balancing algorithm to reduce the impact of local data heterogeneity on model generalization and a server-side passive client selection strategy with a utility function that optimizes both time and accuracy, ensuring that high-quality clients are not excluded. Experiments on CIFAR-10 and EMNIST confirm the method’s effectiveness: our approach improves global accuracy by up to 14.42% and 7.82% in Non-IID settings and achieves up to 1.78 \(\times \) and 1.45 \(\times \) gains in time efficiency compared to baselines.