The source code for ICML2021 paper When Does Data Augmentation Help With Membership Inference Attacks?

Yigitcan Kaya, Tudor Dumitras -- University of Maryland, College Park

Please contact cankaya at umd dot edu for bugs, questions and recommendations.

Requirements:

• CUDA 11.0
• CUDNN 8.0
• PyTorch and TorchVision (pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html)
• Opacus 0.9 - Differential privacy package for PyTorch (pip install opacus==0.9.0)
• tqdm (pip install tqdm)

Source code files and their contents:

• train_models.py --- contains the wrappers to train all the models with and without augmentation, takes the path to a config file as the command line argument.
• mi_attacks.py --- contains the implementations of the three MIAs in the paper (average, powerful and augmentation-aware), take the path to a config file as the command line argument.
• aux_funcs.py --- contains auxiliary functions for optimizers, regularization methods, datasets, loaders etc.
• loss_rank_correlation.py --- contains the implementation of the LRC metric in Section 6.
• models.py --- contains the definitions of the architectures we experiment with, the methods to train models with DP or with augmentation.
• collect_results.py --- contains the methods to collect the results after models are trained and the MIAs are applied.
• Playground.ipynb --- contains the demonstration of how to use the codebase, train models, apply attacks, collect results etc.
• config.json --- the config file for training and attacking the models, specifies the hyper-parameters, paths, and experiment settings.