Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Fair Federated Learning under Demographic Disparities and Data Imbalance

Qiming Wu, Siqi Li, Doudou Zhou, Nan Liu

TL;DR

This work tackles fairness in federated learning under demographic disparities and data imbalance by introducing FedIDA, a framework-agnostic method that jointly addresses algorithmic bias and subgroup underrepresentation. FedIDA combines a convex group-fairness regularizer with a locally applied, fairness-aware ROSE oversampling procedure, preserving convergence with any FL optimizer. Theoretical results establish Lipschitz continuity of fairness metrics and guarantee fairness improvements with high probability, along with variance reduction in fairness across test cohorts. Empirical evaluation on Adult and ROC Epistry datasets demonstrates consistent fairness improvements across multiple disparity metrics with competitive predictive performance, validating FedIDA’s potential for equitable privacy-preserving healthcare modeling. The work also provides practical insights into hyperparameter tuning and ablation studies, while outlining limitations and directions for future work on continuous sensitive attributes and closer integration with FL dynamics.

Abstract

Ensuring fairness is critical when applying artificial intelligence to high-stakes domains such as healthcare, where predictive models trained on imbalanced and demographically skewed data risk exacerbating existing disparities. Federated learning (FL) enables privacy-preserving collaboration across institutions, but remains vulnerable to both algorithmic bias and subgroup imbalance - particularly when multiple sensitive attributes intersect. We propose FedIDA (Fed erated Learning for Imbalance and D isparity A wareness), a framework-agnostic method that combines fairness-aware regularization with group-conditional oversampling. FedIDA supports multiple sensitive attributes and heterogeneous data distributions without altering the convergence behavior of the underlying FL algorithm. We provide theoretical analysis establishing fairness improvement bounds using Lipschitz continuity and concentration inequalities, and show that FedIDA reduces the variance of fairness metrics across test sets. Empirical results on both benchmark and real-world clinical datasets confirm that FedIDA consistently improves fairness while maintaining competitive predictive performance, demonstrating its effectiveness for equitable and privacy-preserving modeling in healthcare. The source code is available on GitHub.

Toward Fair Federated Learning under Demographic Disparities and Data Imbalance

TL;DR

This work tackles fairness in federated learning under demographic disparities and data imbalance by introducing FedIDA, a framework-agnostic method that jointly addresses algorithmic bias and subgroup underrepresentation. FedIDA combines a convex group-fairness regularizer with a locally applied, fairness-aware ROSE oversampling procedure, preserving convergence with any FL optimizer. Theoretical results establish Lipschitz continuity of fairness metrics and guarantee fairness improvements with high probability, along with variance reduction in fairness across test cohorts. Empirical evaluation on Adult and ROC Epistry datasets demonstrates consistent fairness improvements across multiple disparity metrics with competitive predictive performance, validating FedIDA’s potential for equitable privacy-preserving healthcare modeling. The work also provides practical insights into hyperparameter tuning and ablation studies, while outlining limitations and directions for future work on continuous sensitive attributes and closer integration with FL dynamics.

Abstract

Ensuring fairness is critical when applying artificial intelligence to high-stakes domains such as healthcare, where predictive models trained on imbalanced and demographically skewed data risk exacerbating existing disparities. Federated learning (FL) enables privacy-preserving collaboration across institutions, but remains vulnerable to both algorithmic bias and subgroup imbalance - particularly when multiple sensitive attributes intersect. We propose FedIDA (Fed erated Learning for Imbalance and D isparity A wareness), a framework-agnostic method that combines fairness-aware regularization with group-conditional oversampling. FedIDA supports multiple sensitive attributes and heterogeneous data distributions without altering the convergence behavior of the underlying FL algorithm. We provide theoretical analysis establishing fairness improvement bounds using Lipschitz continuity and concentration inequalities, and show that FedIDA reduces the variance of fairness metrics across test sets. Empirical results on both benchmark and real-world clinical datasets confirm that FedIDA consistently improves fairness while maintaining competitive predictive performance, demonstrating its effectiveness for equitable and privacy-preserving modeling in healthcare. The source code is available on GitHub.
Paper Structure (31 sections, 2 theorems, 31 equations, 4 figures, 7 tables, 2 algorithms)

This paper contains 31 sections, 2 theorems, 31 equations, 4 figures, 7 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Algorithmic Disparities in FL.
  4. Data Imbalance in FL.
  5. Joint Fairness and Imbalance.
  6. Method
  7. Preliminaries
  8. Algorithmic Disparities.
  9. Data Imbalance.
  10. Federated Learning.
  11. FedIDA
  12. Theoretical Analysis
  13. Experiments
  14. Datasets
  15. Experiment Setup
  16. ...and 16 more sections

Key Result

Theorem 1

Let $\mathcal{M}(f(\cdot; \mathbf{w}), D)$ be a group fairness metric that is Lipschitz-continuous with respect to group-wise prediction (as in Remark remark:lipschitz), and let $f(\cdot; \mathbf{w}) \in [0,1]$ be fixed and bounded. Suppose dataset $D$ is imbalanced across sensitive-outcome subgroup where $L$ is the Lipschitz constant of $\mathcal{M}$, and $\epsilon > 0$ reflects the expected fair

Figures (4)

  • Figure S1: Results for Logistic Regression on Adult dataset
  • Figure S2: Results for Fully Connected Neural Network on Adult dataset
  • Figure S3: Results for Logistic Regression on ROC dataset
  • Figure S4: Results for Fully Connected Neural Network on ROC dataset

Theorems & Definitions (7)

  • Remark 1: FedIDA Preserves Convergence Properties of $\mathcal{F}$
  • Remark 2: Lipschitz Continuity of Fairness Metrics
  • Theorem 1: Fairness Gain via FairnessAwareROSE
  • Theorem 2: Variance Reduction in Fairness Metrics via FedIDA
  • Remark 3: Lipschitz Continuity of DPR
  • proof
  • proof