HaokunChen245 / FRAug

[ICCV 2023] FRAug: Tackling Federated Learning with Non-IID Features via Representation Augmentation

FRAug (Federated Representation Augmentation)

We proposed a representation augmentation method to improve the model performance in federated learning with feature shift across client local data. [PDF]

---

Setup

See the requirements.txt for environment configuration.

pip install -r requirements.txt

Dataset Preparation

Download the benchmark datasets using the provided script

python datasets/dataset_download.py --dataset="Digits" --data_dir=$PATH_TO_DATASETS$

python datasets/dataset_download.py --dataset="PACS" --data_dir=$PATH_TO_DATASETS$

python datasets/dataset_download.py --dataset="OfficeHome" --data_dir=$PATH_TO_DATASETS$

Training

python main.py --dataset=$DATASET$ --dataset_dir=$PATH_TO_DATASETS$ --log_dir=$PATH_TO_LOG$

Evaluation

python eval.py --dataset=$DATASET$ --model_dir=$PATH_TO_MODEL$ --dataset_dir=$PATH_TO_DATASETS$

---

Citation

@InProceedings{Chen_2023_ICCV,
    author    = {Chen, Haokun and Frikha, Ahmed and Krompass, Denis and Gu, Jindong and Tresp, Volker},
    title     = {FRAug: Tackling Federated Learning with Non-IID Features via Representation Augmentation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {4849-4859}
}