Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Learning and Class Imbalances

Siqi Zhu, Joshua D. Kaggie

TL;DR

The paper addresses federated learning under label noise and class imbalances, proposing a rigorous reproduction and extension of RHFL+ using NVIDIA’s NVFlare framework. It introduces Dynamic Local Noise Learning and Enhanced Client Confidence Reweighting to robustly aggregate heterogeneous client updates under noisy labels, and implements a modular NVFlare-based codebase to enable deployment-ready experimentation. The work extends RHFL+ to medical-imaging datasets (CBIS-DDSM, BreastMNIST, BHI) and conducts extensive ablations, scaling tests, and cross-dataset analyses to validate robustness and scalability. Overall, the study demonstrates RHFL+’s resilience to noisy, non-IID data in heterogeneous FL and provides a practical, reusable framework for future research and real-world deployment in privacy-sensitive domains.

Abstract

Federated Learning (FL) enables collaborative model training across decentralized devices while preserving data privacy. However, real-world FL deployments face critical challenges such as data imbalances, including label noise and non-IID distributions. RHFL+, a state-of-the-art method, was proposed to address these challenges in settings with heterogeneous client models. This work investigates the robustness of RHFL+ under class imbalances through three key contributions: (1) reproduction of RHFL+ along with all benchmark algorithms under a unified evaluation framework; (2) extension of RHFL+ to real-world medical imaging datasets, including CBIS-DDSM, BreastMNIST and BHI; (3) a novel implementation using NVFlare, NVIDIA's production-level federated learning framework, enabling a modular, scalable and deployment-ready codebase. To validate effectiveness, extensive ablation studies, algorithmic comparisons under various noise conditions and scalability experiments across increasing numbers of clients are conducted.

Federated Learning and Class Imbalances

TL;DR

The paper addresses federated learning under label noise and class imbalances, proposing a rigorous reproduction and extension of RHFL+ using NVIDIA’s NVFlare framework. It introduces Dynamic Local Noise Learning and Enhanced Client Confidence Reweighting to robustly aggregate heterogeneous client updates under noisy labels, and implements a modular NVFlare-based codebase to enable deployment-ready experimentation. The work extends RHFL+ to medical-imaging datasets (CBIS-DDSM, BreastMNIST, BHI) and conducts extensive ablations, scaling tests, and cross-dataset analyses to validate robustness and scalability. Overall, the study demonstrates RHFL+’s resilience to noisy, non-IID data in heterogeneous FL and provides a practical, reusable framework for future research and real-world deployment in privacy-sensitive domains.

Abstract

Federated Learning (FL) enables collaborative model training across decentralized devices while preserving data privacy. However, real-world FL deployments face critical challenges such as data imbalances, including label noise and non-IID distributions. RHFL+, a state-of-the-art method, was proposed to address these challenges in settings with heterogeneous client models. This work investigates the robustness of RHFL+ under class imbalances through three key contributions: (1) reproduction of RHFL+ along with all benchmark algorithms under a unified evaluation framework; (2) extension of RHFL+ to real-world medical imaging datasets, including CBIS-DDSM, BreastMNIST and BHI; (3) a novel implementation using NVFlare, NVIDIA's production-level federated learning framework, enabling a modular, scalable and deployment-ready codebase. To validate effectiveness, extensive ablation studies, algorithmic comparisons under various noise conditions and scalability experiments across increasing numbers of clients are conducted.
Paper Structure (55 sections, 11 equations, 18 figures, 9 tables, 3 algorithms)

This paper contains 55 sections, 11 equations, 18 figures, 9 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Why Federated Learning?
  3. Challenges
  4. Class Imbalances
  5. State-of-the-Art Benchmarks
  6. Goal
  7. Structure of Thesis
  8. Overview of Federated Learning
  9. Problem Formulation
  10. Categorization
  11. Homogeneous Federated Learning
  12. Reproduction
  13. Dynamic Local Noise Learning
  14. DLR10816157
  15. SLwang2019symmetriccrossentropyrobust
  16. ...and 40 more sections

Figures (18)

  • Figure 1: NVFlare Learner-Controller Workflow, adapted fromnvflare
  • Figure 2: Data Pipeline of the Experiments Using NVFlare Framework
  • Figure 3: Comparison of all methods performance under different noise levels ($\mu=0.1$ vs $\mu=0.2$)
  • Figure 4: Comparison of all methods performance under different noise levels ($\mu=0.1$ vs $\mu=0.2$)
  • Figure 5: Test accuracy of RHFL+ with increasing number of clients, CIFAR-10 as the Private Dataset, CIFAR-100 as the Public Dataset
  • ...and 13 more figures