FedPD: Federated Open Set Recognition with Parameter Disentanglement

Official implementation of the paper FedPD: Federated Open Set Recognition with Parameter Disentanglement (ICCV 2023).

Setup

cd data
pip install -e .

Dataset & Pretrained Modeel

Benchmark(Digits)

• Please download our pre-processed datasets here, put under data/ directory and perform following commands:

  cd ./data
  unzip digit_dataset.zip

Training

Baseline

Our local training is based on popolar open-set recognition framework Proser. A simple baseline for federated open-set recognition is provided:

python tools/proser_federated.py --log --mode fedavg

FedPD

To achieve parameter alignment for FedOSR, you can try:

python tools/proser_fedpd.py --log --mode fedpd

Visualization of Parameter Misalignment

Dense Open-set Recognition

Acknowledgement

We sincerely thank Proser, DenseHybrid, and FedBN for providing their wonderful code!

Citations

Please use the following bib entry to cite the paper if you are using resources from this repo.

@InProceedings{Yang_2023_ICCV,
    author = {Yang, Chen and Zhu, Meilu and Liu, Yifan and Yuan, Yixuan},
    title = {FedPD: Federated Open Set Recognition with Parameter Disentanglement},
    booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
    year = {2023},
}