This is Pytorch code for our paper Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation. In this paper, we move a step forward and show the existence of a training-free adversarial perturbation under the no-box threat model, which can be successfully used to attack different DNNs in real-time. Extensive experiments on the ImageNet dataset demonstrate the effectiveness of the proposed no-box method. It attacks ten well-known models with a success rate of 98.13% on average, which outperforms state-of-the-art no-box attacks by 29.39%. Furthermore, our method is even competitive to mainstream transfer-based black-box attacks.
- python3.7
- torch 1.5.0
- torchvision 0.6.0
- numpy 1.16.6
- opencv-python 4.5.1.48
- timm 0.4.12
- matplotlib 3.3.4
-
Put the ImageNet dataset into "input/" whose structure like the following:
No-box-HIT-Attack |───input | |───n01440764 | |───n01443537 | |───n01484850 | |───n01491361 | |───n01494475 ..... | |───n15075141
where each folder represents a category and
torchvision.datasets.ImageFolder
can automatically get their corresponding labels. For other dataset, you may need to write a dataloader by yourself. -
Run the below code to perform our HIT attack:
python HIT.py
If you think our work is intersting or useful in your research, please consider citing:
@article{zhang2022hit,
author = {Qilong Zhang and
Chaoning Zhang and
Chaoqun Li and
Jingkuan Song and
Lianli Gao and
Heng Tao Shen},
title = {Practical No-box Adversarial Attacks with Training-free Hybrid Image
Transformation},
journal = {CoRR},
volume = {abs/2203.04607},
year = {2022}
}