This repository provides resources for the following paper:

[Robust Federated Learning with Noisy and Heterogeneous Client](CVPR 2022 Open Access Repository)
Xiuwen Fang, Mang Ye CVPR 2022

RHFL (Robust Heterogeneous Federated Learning) is a federated learning framework to solve the robust federated learning problem with noisy and heterogeneous clients:

1. Aligning the logits output distributions in heterogeneous federated learning.

2. Local noise learning with a noise-tolerant loss function.

3. Client confidence re-weighting for external noise.

Our experiments are conducted on two datasets, Cifar10 and Cifar100. We set public dataset on the server as a subset of Cifar100, and randomly divide Cifar10 to different clients as private datasets.

Dataset used: CIFAR-10、CIFAR-100

Note: Data will be processed in init_data.py

After installing MindSpore via the official website, you can start training and evaluation as follows:

# init public data and local data
python Dataset/init_data.py
# pretrain local models
python Network/pretrain.py
# RHFL
python HHF/HHF.py

├── Robust_FL
    ├── Dataset
        ├── cifar.py
        ├── init_dataset.py
        ├── utils.py
    ├── Network
        ├── Models_Def
            ├── mobilnet_v2.py
            ├── resnet.py
            ├── shufflenet.py
        ├── pretrain.py
    ├── HHF
        ├── HHF.py
    ├── loss.py
    ├── README.md

In the heterogeneous model scenario, we assign four different networks:ResNet10,ResNet12,ShuffleNet,Mobilenetv2

@inproceedings{fang2022robust,
  title={Robust Federated Learning With Noisy and Heterogeneous Clients},
  author={Fang, Xiuwen and Ye, Mang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10072--10081},
  year={2022}
}