inequality

Inequality phenomenon in $L_{\infty}$ adversarial training.

This repository contains the code and models necessary to replicate the results of our paper:

Inequality phenomenon in $l_{\infty}$-adversarial training, and its unrealized threats (ICLR spotlight)
Paper: https://openreview.net/forum?id=4t9q35BxGr

---

Getting Started

Our pretrained model relies on the work Do Adversarially Robust ImageNet Models Transfer Better?

1. Clone our repo.
2. Install dependencies:

conda create -n inequality_test python=3.8  
conda activate inequality_test 
pip install -r requirements.txt

3. Download pretrained model from the link

---

Running Experiments

1. Evaluation inequality degree of pretrained model

python inequality_test.py

2. Inductive noise attack

python noise_eval.py

3. Inductive occludion attack

python occlusion_eval.py

4. Run saliency examples

python saliency_example.py

Please check the augments in each .py, change the attribution method in utils.py

Citation

@inproceedings{duaninequality,
  title={Inequality phenomenon in $ l\_ $\{$$\backslash$infty$\}$ $-adversarial training, and its unrealized threats},
  author={Duan, Ranjie and Chen, YueFeng and Zhu, Yao and Jia, Xiaojun and Zhang, Rong and others},
  booktitle={The Eleventh International Conference on Learning Representations}
}

If you have any further question, please contact [email protected]