Get Fooled for the Right Reason

Official repository for the NeurIPS 2021 paper Get Fooled for the Right Reason: Improving Adversarial Robustness through a Teacher-guided Curriculum Learning Approach

Dependencies

1. Tensorflow 1.14.0
2. Python 3.7

Datasets

CIFAR10: https://www.cs.toronto.edu/~kriz/cifar.html

Models

modelGTP_cifar10: https://www.dropbox.com/sh/29n2lt08ypjdw67/AABSZlD8nTM08E-bcZv1mdkOa?dl=0

Usage

1. Install dependencies with pip install -r requirements.txt. Prefarably, create an anaconda environment.
2. Download and save datasets in datasets/ folder.
3. Download and save model in models/ folder.
4. Run the python eval_attack.py
5. The evaluation results will be stored in attack_log directory.

Note

Using a GPU is highly recommended.

Code overview

• model_new.py: contains code for IGAM model architectures.
• cifar10_input.py provides utility functions and classes for loading the CIFAR10 dataset.
• PGD_attack.py: generates adversarial examples and save them in attacks/.
• run_attack.py: evaluates model on adversarial examples from attacks/.
• config_attack.py: parameters for adversarial example evaluation.
• eval_attack.py: runs FGSM, PGD-5, PGD-10, PGD-20 attacks and logs the results in attack_log directory. However, you can get results for any attack by modifying the num_steps flag in the code.

Acknowledgements

Useful code bases we used in our work:

• https://github.com/MadryLab/cifar10_challenge (for adversarial example generation and evaluation)
• https://github.com/ashafahi/free_adv_train (for model code)