Overcoming Data Imbalance in Federated Learning with Calibration Weighting

Nathaniel Kang, Yugyeong Lim, Jongho Im

Abstract: Federated Learning (FL) is an essential technique for enabling machine learning on decentralized data while preserving privacy. A major challenge in FL is data imbalance, where skewed class distributions across clients lead to biased global models and reduced performance. This paper presents Federated Learning with Feature Calibration and Client Resampling (FL-FCR), an approach tailored to address this challenge. FL-FCR integrates a calibrated loss function during client training and employs calibration-based resampling at the server level, effectively aligning model training with the true data distributions. Extensive experiments on various datasets, including MNIST, FMNIST, CIFAR-10, and CIFAR-100, demonstrate that FL-FCR consistently outperforms existing FL methods in terms of accuracy, particularly in highly skewed data scenarios. These results highlight the potential of calibration techniques in FL, providing a robust solution for effective distributed machine learning.