microsoft / privgan

PrivGAN: Protecting GANs from membership inference attacks at low cost

privGAN

This repository contains the source code for PrivGan - a novel approach for deterring membership inference attacks on GAN generated synthetic medical data.Currently, the repository contains the jupyter notebooks for various datasets. We will be converting the code into a library in the future. Please visit our paper 'PrivGAN: Protecting GANs from membership inference attacks at low cost' ArXiv Link Accepted at PETS 2021

Version information

1. Python 3.7.3
2. Numpy 1.16.2
3. Pandas 0.25.3
4. Tqdm 4.38.0
5. Keras 2.2.4
6. Scipy 1.1.0
7. Tensorflow 1.14.0
8. Scikit-learn 0.20.3

Notebooks comparing white-box attack accuracy of privGAN and GAN on verious datasets

1. PrivGAN_mnist.ipynb
2. PrivGAN_mnist_fashion.ipynb
3. PrivGAN_lfw.ipynb
4. PrivGAN_cifar.ipynb

Notebooks comparing performance on downstream classification tasks

1. MNIST_down.ipynb

Installation

Contribution

Please review the link here to know code of conduct https://opensource.microsoft.com/codeofconduct . Before submitting a pull request please remove all output from your notebooks by going to Cell -> All Output -> Clear

Contact

Copyright

Copyright (c) Microsoft Corporation.