Private-Set

This repository contains the implementation for "Private Set Generation with Discriminative Information (NeurIPS 2022)".

Contact: Dingfan Chen (dingfan.chen@cispa.de)

Requirements

• This implementation is based on PyTorch (tested for version 1.7.1).
• Please refer to requirements.txt for the other required packages and version.
• The differentially private training part is based on Opacus with version 0.15.0.
  Note that due to significant refactoring, later versions of Opacus (>=1.0.0) is not compatible with our implementation at the moment.

Running Experiments

Classification (without prior)

python main.py \
--dataset "Dataset name" \
--exp_name "Experiment name" (name of the output directory) \
--spc "Number of samples per-class" \
--enable_privacy \
--target_epsilon "The target privacy level (default=10)"

• See main.py or run python main.py --help for a detailed descriptions of all supported arguments.
• Privacy cost computation is done by: automatically computing the required noise_multiplier given the desired target_epsilon.
• Dataset: 'MNIST','FashionMNIST', 'CIFAR10' has been tested.
• Number of samples per-class: spc in [1, 10, 20, 50] has been tested.

Classification (with DCGAN network as image prior)

python main_prior.py \
--dataset "Dataset name" \
--exp_name "Experiment name" (name of the output directory) \
--spc "Number of samples per-class" \
--enable_privacy \
--target_epsilon "The target privacy level (default=10)"

• See main_prior.py or run python main_prior.py --help for a detailed descriptions of all supported arguments.
• Privacy cost computation is done by: automatically computing the required noise_multiplier given the desired target_epsilon.
• Dataset: 'MNIST','FashionMNIST' has been tested.
• Number of samples per-class: spc in [1, 10, 20] has been tested.

Continual (Class-incremental) learning

cd continual_learning 

• Dataset: 'MNIST','FashionMNIST' are supported.
• Run Ours

  python main_ours.py \
  --dataset "Dataset name" \
  --exp_name "Experiment name" (name of the output directory) \
  --enable_privacy \
  --target_epsilon "The target privacy level (default=10)"

• Run DP-SGD baseline

  python main_dpsgd.py \
  --dataset "Dataset name" \
  --exp_name "Experiment name" (name of the output directory) \
  --enable_privacy \
  --target_epsilon "The target privacy level (default=10)"

• Run DP-Merf baseline

  python main_dpmerf.py \
  --dataset "Dataset name" \
  --exp_name "Experiment name" (name of the output directory) \
  --enable_privacy \
  --target_epsilon "The target privacy level (default=10)"

Citation

@inproceedings{chen2022privateset,
  title={Private Set Generation with Discriminative Information},
  author={Chen, Dingfan and Kerkouche, Raouf and Fritz, Mario},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2022}
}

Acknowledgements

Our implementation uses the source code from the following repositories:

• Dataset Condensation
• DCGAN prior from GS-WGAN
• Class incremental learning
• DP-Merf baseline