Reproduce “Multi-Level Analyzation of Imbalance to Resolve Non-IID-Ness in Federated Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haengbok Chung, Jae Sung Lee

Researchers from Seoul National University introduce FedBB, a Federated Learning framework designed to mitigate performance drops caused by non-IID data through a multi-level analysis of class imbalance. It employs a Positive Negative Balanced (PNB) loss for robust local model training and Client Balanced Reweighting (CBR) for optimized global aggregation, achieving state-of-the-art AUCs on multi-label medical image classification and top accuracies on multi-class natural image datasets under various non-IID conditions.

Key insights

  • Effectively resolving non-IID challenges in Federated Learning requires a comprehensive understanding and mitigation of class imbalance across multiple levels, from individual case distribution to client-level data skewness.
  • Optimizing both local model learning with an imbalance-aware loss function and global model aggregation with a balance-based reweighting strategy is crucial for robust FL performance on heterogeneous data.
  • Superior performance and efficiency in Federated Learning can be achieved by carefully managing class imbalance without incurring prohibitive computational or resource overhead.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.