lq1029 / FedBVA

Adversarial Robustness through Bias Variance Decomposition: A New Perspective for Federated Learning

FedBVA

This is the official implementation of "Adversarial Robustness through Bias Variance Decomposition: A New Perspective for Federated Learning" (CIKM'22).

Requirements

• Python 3.6
• numpy==1.18.1
• torch==1.4.0
• torchvision==0.5.0

Training and testing

For robust federated learning on MNIST image data set with IID setting, please run

python main.py

Details

• utils: hyper-parameter setting (params.py) and decentralized data sampling (sampling.py)
• models: our decentralized learning algorithm, including client update and server update
• data: image data set (e.g., MNIST, Fashion-MNIST, Cifar10 and Cifar100) could be downloaded automatically from torchvision package
• save: we use the same random model initialization for all the experiments, blackbox attack models need to be pretrained

Acknowledgement

This is the latest source code of FedBVA. If you find that it is helpful for your research, please consider to cite our paper:

@inproceedings{zhou2020adversarial,
  title={Adversarial Robustness through Bias Variance Decomposition: A New Perspective for Federated Learning},
  author={Zhou, Yao and Wu, Jun and Wang, Haixun and He, Jingrui},
  booktitle={Proceedings of the 31th ACM International Conference on Information and Knowledge Management},
  year={2022}
}