Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss

This repository is the offical implementation for the paper: Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss (ICML2024)

Installation

cd ConvexConcaveLoss;
conda env create -f environment.yml;
conda activate ccl;
python setup.py install;

Membership inference attacks

Step 1: train target/shadow models

cd ConvexConcaveLoss/source/examples;
python train_models.py --mode target --training_type Normal --loss_type ccel --alpha 0.5 --beta 0.05 --gpu 0 --optimizer sgd --scheduler multi_step --epoch 300 --learning_rate 0.1;
python train_models.py --mode shadow --training_type Normal --loss_type ccel --alpha 0.5 --beta 0.05 --gpu 0 --optimizer sgd --scheduler multi_step --epoch 300 --learning_rate 0.1;

Note that you can also specify the --loss_type with different loss function, e.g., ce, focal and ccql.

Step 2: perform membership inference attacks

python mia.py  --training_type Normal --loss_type ccel --attack_type metric-based --alpha 0.5 --beta 0.05 --gpu 0 --scheduler multi_step --epoch 300 --learning_rate 0.1;

Citation

If you find this useful in your research, please consider citing:

@inproceedings{liu2024mitigating,
  title={Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss},
  author={Zhenlong Liu and Lei Feng and Huiping Zhuang and Xiaofeng Cao and Hongxin Wei},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2024}
}

Acknowledgements

Our implementation uses the source code from the following repositories: MLHospital