Interpreting Adversarially Trained Convolutional Neural Networks

Code for our ICML'19 paper: Interpreting Adversarially Trained Convolutional Neural Networks, by Tianyuan Zhang, Zhanxing Zhu.

How to run

If you want to run code on your own, I only provide codes on Caltech-256. We do not upload the dataset and trainned networks due to the fact that they are storage consuming. Training(Adversarial Training) scripts in this repo is not well writen, I suugest you to use your own scripts, or scripts provided in this repo

Caltech-256

The four most important python files for Caltech-256

They are all in /code/baseline/ :

main.py, attack.py, utils.py dataset.py

1. main.py trains CNNs. It contains sufficient comments to understand how to customize your trainings.
2. attack.py implements PGD and FGSM attackers.
3. utils.py implements SmoothGrad
4. dataset.py implements Saturation and Patch-shuffle operation. For style transfer, you have to use code provided in https://github.com/rgeirhos/Stylized-ImageNet/tree/master/code
5. stAdv.py in code/spatial-transform.adv.train/ implements Spatially transformed attack.

How to run

1. Dowload the data from http://www.vision.caltech.edu/Image_Datasets/Caltech101/Caltech101.html to /code/Caltech256/data
2. cd data and run Partition.py to generate training set and test set.
3. cd code/baseline and run main.py to train standard CNNs
4. cd code/pgd.inf.eps8 and run main.py to adversarially train CNNs against a $l_{\inf}$ -norm bounded PGD attacker. code in /code/Caltech256/code/pgd.l2.eps8 are for $l_2$ -norm bounded adversarial training .
5. gen_visual.py and utils.py in /code/Caltech256/code/baseline/ contains code to generate salience maps.

TinyImageNet & CIFAR-10

Code for TinyImageNet, and CIAFR-10 are similar to that for Caltech-256, important python files such as main.py, utils.py, attack.py has the same name and functionallities

You still have to download the images from https://tiny-imagenet.herokuapp.com and put them to /code/TinyImageNet/data

Reference

@article{zhang2019interpreting,
  title={Interpreting Adversarially Trained Convolutional Neural Networks},
  author={Zhang, Tianyuan and Zhu, Zhanxing},
  journal={arXiv preprint arXiv:1905.09797},
  year={2019}
}

Please contact tianyuanzhang if you get any question on this paper or the code. ) （＾∀＾）