A privacy-preserving, group-fair federated learning framework that mitigates bias arising from feature heterogeneity across edge devices by leveraging multiplicative-weights updates with regularization and performance thresholds.
- Introduction
- Installation & Prerequisites
- Data Preparation
- Configuration
- Usage
- Experiments & Results
- Citation
Federated learning enables decentralized model training on heterogeneous edge devices without sharing raw data, preserving user privacy while aggregating local updates into a global model. However, variation in device quality and sensing environments introduces feature heterogeneity, causing global models to underperform on groups with noisier data and perpetuate bias.
We introduce a Multiplicative Weights update with Regularization (MWR) framework that:
- Computes privacy-preserving group importance weights via average conditional probabilities across clients.
- Applies a modified multiplicative-weights algorithm with an L1 regularizer to prevent weight explosion and improve worst-group performance.
- Enforces a performance threshold on the best-performing group to avoid degrading high-accuracy groups.
- Demonstrates superior worst-group true-positive-rate gains (up to +41%) on CIFAR-10, MNIST, Fashion-MNIST, USPS, SynthDigits, and MNIST-M without significant loss in overall accuracy.
- Clone the repository
git clone https://github.com/emtechlab/mitigating-group-bias-in-fl.git cd mitigating-group-bias-in-fl - Create and activate a virtual environment
python3 -m venv venv source venv/bin/activate # on Linux/macOS
- Install required packages
pip install -r requirements.txt
- Training
cd src/ ./run.sh
The training outputs—model checkpoints, logs, and evaluation metrics—will be saved in the evaluation/ directory at the project root (e.g. ../evaluation/).