FedReLa re-labels imbalanced data to boost federated learning accuracy

Federated learning (FL) trains models across decentralized data while preserving privacy, but real-world data is often imbalanced and heterogeneous. A new paper on arXiv introduces FedReLa, a data-level approach that re-labels samples using a feature-dependent label re-allocator. This corrects biased global decision boundaries without needing the global class distribution—a key limitation of prior methods. By relabeling misaligned samples, FedReLa improves minority class accuracy and overall performance on stepwise-imbalanced and long-tailed datasets, surpassing previous state-of-the-art methods without extra communication costs.

The method is modular, model-agnostic, and can be integrated with existing algorithmic FL approaches. The authors tested on multiple benchmarks, showing consistent improvements even under extreme class absence across clients. This work addresses a critical pain point in FL: the coexistence of data heterogeneity and class imbalance, which degrades model performance. FedReLa's re-labeling strategy is a practical solution for enterprises deploying FL in non-IID environments, such as healthcare or finance, where skewed class distributions are common.