Awesome

Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains (ICLR'2022)

This is the Pytorch code for our paper Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains). In this paper, with only the knowledge of the ImageNet domain, we propose a Beyond ImageNet Attack (BIA) to investigate the transferability towards black-box domains (unknown classification tasks).

Requirement

• Python 3.7
• Pytorch 1.8.0
• torchvision 0.9.0
• numpy 1.20.2
• scipy 1.7.0
• pandas 1.3.0
• opencv-python 4.5.2.54
• joblib 0.14.1
• Pillow 6.1
• pretrainedmodels 0.7.4

Dataset

• Download the ImageNet training dataset.

  • ImageNet Training Set.

• Download the testing dataset.

  • ImageNet Validation Set. We also provide precomputed imagenet validation dataset in here (download it to imagenet folder and then run convert.py, thanks to @aaron-xichen).
  • CUB-200-2011
  • Stanford Cars
  • FGVC AirCraft
  • CIFAR-10, CIFAR-100, STL-10 and SVHN can be automatically downloaded via torchvision.dataset

Note: After downloading CUB-200-2011, Standford Cars and FGVC Aircraft, you should set the "self.rawdata_root" (DCL_finegrained/config.py: lines 59-75) to your saved path.

Target model

The checkpoint of target model should be put into model folder.

• CUB-200-2011, Stanford Cars and FGVC AirCraft can be downloaded from here.
• CIFAR-10, CIFAR-100, STL-10 and SVHN can be automatically downloaded.
• ImageNet pre-trained models are available at torchvision.

Pretrained-Generators

Adversarial generators are trained against following four ImageNet pre-trained models.

• VGG19
• VGG16
• ResNet152
• DenseNet169

After finishing training, the resulting generator will be put into saved_models folder. You can also download our pretrained-generator from here.

Train

Train the generator using vanilla BIA (RN: False, DA: False) against ImageNet pretrained VGG-16

python train.py --model_type vgg16 --train_dir your_imagenet_path --RN False --DA False

your_imagenet_path is the path where you download the imagenet training set.

Evaluation

Evaluate the performance of vanilla BIA (RN: False, DA: False)

python eval.py --model_type vgg16 --RN False --DA False

Citing this work

If you find this work is useful in your research, please consider citing:

@inproceedings{Zhang2022BIA,
  author    = {Qilong Zhang and
               Xiaodan Li and
               Yuefeng Chen and
               Jingkuan Song and
               Lianli Gao and
               Yuan He and
               Hui Xue},
  title     = {Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains},
  Booktitle = {International Conference on Learning Representations},
  year      = {2022}
}

Acknowledge

Thank @aaron-xichen, @Muzammal-Naseer and @JDAI-CV for sharing their codes.