[ICLR25] On the Byzantine-Resilience of Distillation-Based Federated Learning

Authors: Christophe Roux, Max Zimmer, Sebastian Pokutta

This repository contains the code to reproduce the experiments from the ICLR25 paper "On the Byzantine-Resilience of Distillation-Based Federated Learning". The code is based on PyTorch 1.9 and the experiment-tracking platform Weights & Biases.

Structure and Usage

Structure

Experiments are started from the following file:

• main.py: Starts experiments using the dictionary format of Weights & Biases.

The rest of the project is structured as follows:

• byzantine: Contains the attacks and defenses used in the paper.
• runners: Contains classes to control the training and collection of metrics.
• models: Contains all model architectures used.
• utilities.py: Contains useful auxiliary functions and classes.
• config.py: Configuration for the datasets used in the experiments.
• public_config.py: Contains the configuration for the public datasets.
• metrics.py: Contains the metrics used in the experiments.
• strategies.py: Contains the different strategies used, such as FedAVG and FedDistill.

Usage

Define the parameters in the main.py defaults-dictionary and run it with the --debug flag. Or, configure a sweep in Weights & Biases and run it from there (without the flag).

Citation

In case you find the paper or the implementation useful for your own research, please consider citing:

@inproceedings{roux2025on,
title={On the Byzantine-Resilience of Distillation-Based Federated Learning},
author={Christophe Roux and Max Zimmer and Sebastian Pokutta},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=of6EuHT7de}
}