Hao Jiang; Jintao Yang; Guang Hua*; Lixia Li; Ying Wang; Shenghui Tu; Song Xia

[*: corresponding author]

This repository contains the implementation of our paper FAWA: Fast Adversarial Watermark Attack

If you find this code or the paper useful, please consider citing:

@ARTICLE{fawa,
  author={Jiang, Hao and Yang, Jintao and Hua, Guang and Li, Lixia and Wang, Ying and Tu, Shenghui and Xia, Song},
  journal={IEEE Transactions on Computers}, 
  title={FAWA: Fast Adversarial Watermark Attack}, 
  year={2021},
  volume={},
  number={},
  pages={1-13},
  doi={10.1109/TC.2021.3065172}}

1. Installation
2. Usage
   • Generate fawa examples
   • Evaluate fawa examples
   • Visualization (CAM)

1. Clone this repository:

git clone https://github.com/JintaoYang18/FAWA
cd FAWA/

2. Install conda envs and requirements:

conda env create -f fawa_env.yaml
pip install -r fawa_requirements.txt

Note: if you don't have a GPU, install the cpu version of PyTorch. (We have not tested this setting.)

3. Prepare your dataset and put it into 100_image_class_950_999_300resize directory.

4. Modify the 3 .txt files in the root directory according to your own data.

Note: We explained the role of each .txt file in detail in the main.py file.

5. Create pre_trained_models directories and download pre-trained .pth files:

mkdir pre_trained_models
cd pre_trained_models/

Download vgg-16 pre-trained .pth file, and put it in the pre_trained_models directory.

python main.py

Note: The time it takes to generate the image depends on the performance of the computer, and you may have to wait. The running time can be adjusted by modifying p_size and g_round. You can even reduce the dimension of the problem, such as removing the rotation item.

Note: Use your own model or pre-trained model to evaluate fawa adversarial examples.

Note: You can use open source code, such as:

• https://github.com/utkuozbulak/pytorch-cnn-visualizations

• https://github.com/jacobgil/pytorch-grad-cam