thalesgroup / trustgan

Training safe and trustworthy deep learning models through generative adversarial networks

TrustGAN

TrustGAN: Training safe and trustworthy deep learning models through generative adversarial networks

This package provides the code developped for the paper:
"TrustGAN: Training safe and trustworthy deep learning models through generative adversarial networks"
presented at the CAID-2022 (Conference on Artificial Intelligence for Defence) https://arxiv.org/abs/2211.13991

Install

With python, pip and setuptools installed, you simply need to:

python -m pip install .

Get the data and train a model

We will work within ./xps/:

Get the datasets

You can download in-distribution (ID) sample datasets within data/:

python3 ../bin/trustgan-download-data.py --path2save "data/" --dataset "MNIST"

You can download out-of-distribution (OOD) sample datasets within data/:

python3 ../bin/trustgan-download-data.py --path2save "data/" --dataset "FashionMNIST"
python3 ../bin/trustgan-download-data.py --path2save "data/" --dataset "CIFAR10"

Train a model

We will now train two models, one without TrustGAN and another with it, with a selected device <device>:

python3 ../bin/trustgan-model-gan-combined-training.py \
    --path2save "mnist-wo-gan/" \
    --path2dataset "data/MNIST" \
    --nb-classes 10 \
    --prop-net-alone 1 \
    --num-epochs 100 \
    --batch-size 512 \
    --device "cuda:0"

python3 ../bin/trustgan-model-gan-combined-training.py \
    --path2save "mnist-wi-gan/" \
    --path2dataset "data/MNIST" \
    --nb-classes 10 \
    --num-epochs 100 \
    --batch-size 512 \
    --nb-step-net-alone 1 \
    --device "cuda:1"

Test

You can get summary plots and gifs with:

python3 ../bin/trustgan-model-gan-combined-training.py \
    --path2save "mnist-wo-gan/" \
    --path2dataset "data/MNIST" \
    --nb-classes 10 \
    --produce-plots

python3 ../bin/trustgan-model-gan-combined-training.py \
    --path2save "mnist-wi-gan/" \
    --path2dataset "data/MNIST" \
    --nb-classes 10 \
    --produce-plots

Contributing

If you are interested in contributing to the project, start by reading the Contributing guide.

License

This repository is licensed under the terms of the MIT License (see the file LICENSE).

Citing

Please cite the following paper if you are using TrustGAN

@ARTICLE{2022arXiv221113991D,
       author = {{du Mas des Bourboux}, H{\'e}lion},
        title = "{TrustGAN: Training safe and trustworthy deep learning models through generative adversarial networks}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Machine Learning, Computer Science - Computer Vision and Pattern Recognition},
         year = 2022,
        month = nov,
          eid = {arXiv:2211.13991},
        pages = {arXiv:2211.13991},
archivePrefix = {arXiv},
       eprint = {2211.13991},
 primaryClass = {cs.LG},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2022arXiv221113991D},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}