[2025.01.22] - 🎊 PRISM has been accepted by ICLR 2025! 🎊
conda create -n prism python=3.10.9
conda activate prism
pip install -r requirements.txt
pip3 install torch torchvision torchaudio
mkdir embeddings embeddings/npz embeddings/precompute
wandb login PERSONAL_API_KEY
- --save=True is an optional, only when logging to WANDB.
python ./run_prism.py --model=prism --aggregation=BA --dataset={mnist|fmnist|celeba|cifar} --gpu=0 --iid=1 --gpunum=0 --num_scorelayer=0 --epochs=500 --experiments=EXRIMENTS --dp_epsilon=9.8 --MADA
- Shards split : --split shards --divide=4
- dirichlet split : --split dirichlet --dir_alpha 0.005
python ./run_prism.py --model=prism --aggregation=BA --dataset={mnist|fmnist|celeba|cifar} --gpu=0 --iid=0 --split {shards | dirichlet} --divide=4 --dir_alpha 0.005 --gpunum=0 --num_scorelayer=0 --epochs=500 --experiments=EXRIMENTS --dp_epsilon=9.8 --MADA
- --num_scorelayer=0~1, for PRISM-$\alpha$. Default configuration is 0 (equivalent with PRISM).
@article{seo2025prism,
title={PRISM: Privacy-Preserving Improved Stochastic Masking for Federated Generative Models},
author={Seo, Kyeongkook and Han, Dong-Jun and Yoo, Jaejun},
journal={arXiv preprint arXiv:2503.08085},
year={2025}
}