The codebase is an academic research prototype, and meant to elucidate protocol details and for proofs-of-concept, and benchmarking. It is not meant for deployment currently.

F2ED-LEARNING: Attacks and Byzantine-Robust Aggregators in Federated Learning

This repository contains the evaluation code for the following manuscripts.

Byzantine-Robust Federated Learning with Optimal Statistical Rates and Privacy Guarantees. Banghua Zhu*, Lun Wang*, Qi Pang*, Shuai Wang, Jiantao Jiao, Dawn Song, Michael Jordan.
Towards Bidirectional Protection in Federated Learning. Lun Wang*, Qi Pang*, Shuai Wang, Dawn Song. SpicyFL Workshop @ NeurIPS 2020.

Attacks

We implemented the following attacks in federated learning.

Byzantine-Robust Aggregators

We implemented the following Byzantine-robust aggregators in federated learning.

Dependency

conda 4.12.0
Python 3.7.11
Screen version 4.06.02 (GNU) 23-Oct-17

First, create a conda virtual environment with Python 3.7.11 and activate the environment.

conda create -n venv python=3.7.11
conda activate venv

Run the following command to install all the required python packages.

pip install -r requirements.txt

Usage

Reproduce the evaluation results by running the following script. You might want to change the GPU index in the script manually. The current script distributes training tasks to 8 Nvidia GPUs indexed by 0-7.

./train.sh

To run a single Byzantine-robust aggregator against a single attack on a dataset, run the following command with the right system arguments:

python src/simulate.py --dataset='dataset' --attack='attack' --agg='aggregator'

For DBA attack, we reuse its official implementation. First open a terminal and run the following command to start Visdom monitor:

python -m visdom.server -p 8097

Then start the training with selected aggregator and attack, which are specified in utils/X.yaml, X can be mnist_params or fashion_params.

cd ./src/DBA
python main.py --params utils/X.yaml

For GAN aggregator, run the following command to start training in round X:

python src/simulate_gan.py --current_round=X --attack='noattack' --dataset='MNIST'
python src/gan.py --next_round=X+1 --gan_lr=1e-5

Note that X starts from 0, and you may try different hyper-parameters like learning rate in gan.py if you use datasets other than MNIST or attacks other than trimmedmean and noattack.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{zhu2022byzantine,
  title={Byzantine-Robust Federated Learning with Optimal Statistical Rates and Privacy Guarantees},
  author={Banghua Zhu and Lun Wang and Qi Pang and Shuai Wang and Jiantao Jiao and Dawn Song and Michael Jordan},
  year={2022},
  url={https://arxiv.org/abs/2205.11765}
}

@article{wang2020f,
  title={F2ED-LEARNING: Good fences make good neighbors},
  author={Lun Wang and Qi Pang and Shuai Wang and Dawn Song},
  journal={CoRR},
  year={2020},
  url={http://128.1.38.43/wp-content/uploads/2020/12/Lun-Wang-07-paper-Lun.pdf}
}

Acknowledgement

The code of evaluation on DBA attacks largely reuse the original implementation from the authors of DBA.

wanglun1996/secure-robust-federated-learning