P-ALPhA

Overview

The PyTorch implementation of our paper:

Chenchen Zhao, and Hao Li. Condition-Invariant Physical Adversarial Attacks via Pixel-Wise Adversarial Learning. International Conference on Neural Information Processing 2021
[paper] [learn more]

We propose a physical adversarial attack method that maintains high robustness against multiple simulated real-world disturbances (e.g. imbalanced illumination, long distances, noise in capturing, etc). The model reaches state-of-the-art physical attack performance, and is the first one to propose the concept of pixel-wise adversarial learning in adversarial attacks

1.5 years later...

• Rewrite the code

Usage

Run conda env create -f environment.yaml && conda activate palpha to create and activate a conda virtual environment named palpha

Run python main.py attack to conduct P-ALPhA on a pretrained classifier

Run python main.py test to validate the performance of the model against different simulated real-world disturbances

Modify args.py for customized experimental settings

Modify model.py for customized classifier

This project does not include classifier-training scripts. Pre-train a classifier, and replace model.py with customized classifier definition. The classifier should only have logits as output

Performance

ResNet18	Success rate %	Illumination %	Distances %	Errors %
PGD & P-ALPhA	100.0	98.0	100.0	91.8
PGD	100.0	6.07	4.49	3.34
EoT 1	16.0	40.5	87.3	77.9
C&W & P-ALPhA	100.0	99.0	79.6	59.2
C&W	100.0	92.9	75.5	30.6

VGG16	Success rate %	Illumination %	Distances %	Errors %
PGD & P-ALPhA	99.0	95.9	100.0	74.5
PGD	100.0	8.08	6.06	12.1
EoT 1	50.5	12.0	54.0	12.0
C&W & P-ALPhA	100.0	39.4	53.5	11.1
C&W	100.0	33.3	53.5	13.1

BibTeX

@inproceedings{zhao2021condition,
  title={Condition-Invariant Physical Adversarial Attacks via Pixel-Wise Adversarial Learning},
  author={Zhao, Chenchen and Li, Hao},
  booktitle={International Conference on Neural Information Processing},
  pages={369--380},
  year={2021},
  organization={Springer}
}

References

[1] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International Conference on Machine Learning, pages 284-293, 2018.