Multi-Level Analyzation of Imbalance to Resolve Non-IID-Ness in Federated Learning

Researchers propose FedBB, a federated learning framework that addresses class imbalance across three levels—within classes, between classes, and across distributed clients—using a specialized loss function and client reweighting strategy. The approach improves model performance on non-IID data while minimizing privacy risks through limited statistical information requirements.

Federated learning enables training machine learning models across decentralized data sources without centralizing sensitive information, making it critical for privacy-preserving applications in healthcare, finance, and other regulated industries. However, real-world federated environments suffer from non-identically distributed (non-IID) data, where each client's local dataset has different class distributions. This heterogeneity degrades model performance because global models trained on imbalanced data struggle to generalize across all classes.

The FedBB framework tackles this through two innovations. The Positive Negative Balanced loss function reweights training samples to address imbalance at the local level, ensuring minority classes and underrepresented examples receive appropriate attention during training. The Client Balanced Reweighting strategy then adjusts aggregation weights during the global update phase, prioritizing contributions from clients with more balanced datasets. This dual-layer approach addresses a fundamental tension in federated learning: clients cannot share raw data for privacy reasons, yet the system must account for their data characteristics when aggregating models.

The framework's efficiency gains and minimal privacy footprint matter for practical deployment. By requiring only limited statistical information rather than full dataset metadata, FedBB reduces the privacy attack surface while maintaining performance improvements validated on medical imaging and natural image datasets.

For the machine learning community, this represents progress on a persistent technical challenge affecting production federated systems. As organizations increasingly adopt federated architectures for sensitive applications, robust handling of heterogeneous data distributions becomes table-stakes infrastructure rather than a research problem.

• →FedBB addresses class imbalance at three interconnected levels: individual cases, class distribution, and client-level data skewness
• →The framework combines local loss reweighting with client-aware global aggregation to improve model generalization on non-IID data
• →Limited statistical requirements enhance privacy protection compared to methods requiring detailed dataset metadata
• →Experimental validation on medical imaging and natural images demonstrates consistent performance improvements over existing approaches
• →The solution serves as a foundation for both generic and personalized federated learning applications