wzhzzz1 / privatefl

PrivateFL: Accurate, Differentially Private Federated Learning via Personalized Data Transformation

This is the Pytorch implementation of our paper, PrivateFL: Accurate, Differentially Private Federated Learning via Personalized Data Transformation.

Experiment Setup

First enter the following path:

cd script

We use miniconda to create a virtual environment with python 3.8, you can install miniconda use the following script if you are using Linux-x86-64bit machine:

(Optional for install miniconda)

bash install_conda.sh

Then use the following script to download the requirements:

bash setup.sh

Code Usage

Train from scratch

You can use the following script to train from scratch.

bash fedavg.sh

You can also change the parameters in script/train.sh, e.g., --data --nclient --nclass --ncpc --model --mode --round --epsilon --sr --lr, following the choices listed in parse_arguments() of FedAverage.py. The value of the parameters can be found in our paper.

Train with frozen encoder

Run the following script to extract features from [ResNeXt, SimCLR, CLIP] and train a one-layer classifier, you may need to download ResNext using this link:

bash fedtransfer.sh

Please reduce the value of --physical_bs if facing CUDA out of memory.

Citation

@inproceedings{yangprivatefl,
  title={PrivateFL: Accurate, Differentially Private Federated Learning via Personalized Data Transformation},
  author={Yang, Yuchen and Hui, Bo and Yuan, Haolin and Gong, Neil and Cao, Yinzhi}
  booktitle = {Proceedings of the USENIX Security Symposium (Usenix'23)},
  year = {2023}
}