is a Python toolbox for adversarial robustness research. The primary functionalities are implemented in PyTorch. Specifically, AdverTorch contains modules for generating adversarial perturbations and defending against adversarial examples, also scripts for adversarial training.
We developed AdverTorch under Python 3.6 and PyTorch 1.0.0 & 0.4.1. To install AdverTorch, simply run
pip install advertorch
or clone the repo and run
python setup.py install
To install the package in "editable" mode:
pip install -e .
Some attacks are tested against implementations in Foolbox or CleverHans to ensure correctness. Currently, they are tested under the following versions of related libraries.
conda install -c anaconda tensorflow-gpu==1.11.0
pip install git+https://github.com/tensorflow/cleverhans.git@336b9f4ed95dccc7f0d12d338c2038c53786ab70
pip install Keras==2.2.2
pip install foolbox==1.3.2
# prepare your pytorch model as "model"
# prepare a batch of data and label as "cln_data" and "true_label"
# ...
from advertorch.attacks import LinfPGDAttack
adversary = LinfPGDAttack(
model, loss_fn=nn.CrossEntropyLoss(reduction="sum"), eps=0.3,
nb_iter=40, eps_iter=0.01, rand_init=True, clip_min=0.0, clip_max=1.0,
targeted=False)
adv_untargeted = adversary.perturb(cln_data, true_label)
target = torch.ones_like(true_label) * 3
adversary.targeted = True
adv_targeted = adversary.perturb(cln_data, target)For runnable examples see advertorch_examples/tutorial_attack_defense_bpda_mnist.ipynb for how to attack and defend; see advertorch_examples/tutorial_train_mnist.py for how to adversarially train a robust model on MNIST.
The documentation webpage is on readthedocs https://advertorch.readthedocs.io.
AdverTorch is still under active development. We will add the following features/items down the road:
- more examples
- support for other machine learning frameworks, e.g. TensorFlow
- more attacks, defenses and other related functionalities
- support for other Python versions and future PyTorch versions
- contributing guidelines
- ...
FastFeatureAttack and JacobianSaliencyMapAttack do not pass the tests against the version of CleverHans used. (They use to pass tests on a previous version of CleverHans.) This issue is being investigated. In the file test_attacks_on_cleverhans.py, they are marked as "skipped" in pytest tests.
This project is licensed under the LGPL. The terms and conditions can be found in the LICENSE and LICENSE.GPL files.
If you use AdverTorch in your research, we kindly ask that you cite the following technical report:
@article{ding2019advertorch,
title={{AdverTorch} v0.1: An Adversarial Robustness Toolbox based on PyTorch},
author={Ding, Gavin Weiguang and Wang, Luyu and Jin, Xiaomeng},
journal={arXiv preprint arXiv:1902.07623},
year={2019}
}
- Gavin Weiguang Ding
- Luyu Wang
- Xiaomeng Jin
- Laurent Meunier
- Alexandre Araujo
- Jérôme Rony
- Ben Feinstein
- Francesco Croce
- Taro Kiritani