README.md

Demystifying Local & Global Fairness Trade-offs in Federated Learning

This repository contains the code accompanying the paper "Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition" by Faisal Hamman and Sanghamitra Dutta, presented at ICLR 2024.

Our research investigates the trade-offs between local and global fairness in federated learning environments through the lens of Partial Information Decomposition (PID).

Experiments

Experiment A: Accuracy-Global-Local-Fairness Trade-off Pareto Front

• Objective: To study the trade-offs between model accuracy and different fairness constraints.
• Files: Run Adult-tradeoff.ipynb for the Adult dataset and Synthetic-tradeoff.ipynb for synthetic data analysis.

Experiment B: Demonstrating Disparities in Federated Learning Settings

• Objective: Investigate the PID of disparities on the Adult dataset trained within a federated learning (FL) framework using the FedAvg algorithm (McMahan et al., 2017).
• Files: Code implementations are located in the FedAvg directory.

To run the experiments, ensure your environment is properly set up by installing the required packages:

conda env create -f environment.yml

Ensure your Federated Learnining environment is configured correctly by referring to config.yaml for detailed settings.

Run for experiment B:

python main.py

Acknowledgments

The implementation of the FedAvg algorithm in this project was adapted from Federated Learning in PyTorch and Federated Averaging PyTorch. The trade-off analysis code was adapted from FACT.

Please consider citing our work:

@inproceedings{
hamman2024demystifying,
title={Demystifying Local \& Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition},
author={Faisal Hamman and Sanghamitra Dutta},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=SBj2Qdhgew}
}