This project is for the ACDM we proposed in the paper "Artwork Protection Against Neural Style Transfer Using Locally Adaptive Adversarial Color Attack", which is accepted by 50th European Conference on Artificial Intelligence (ECAI 2024).

The eval.py contains the settings for reproducing the result of Table 1 in the main paper.

We use the EFDM (the pre-trained vgg and decoder can be found in the link) to generate the image pairs which have:

1. same content but different style
2. same style but different content

For datasets, we use the COCO and WikiArt (painter by numbers). As some images are too huge to be processed by PIL, so I filtered some files, the files we kept for the project are listed in the style.csv and content.csv.

We used 10,000 pairs of images with random_seed=3407, the result is:

lpips==0.1.4
numpy==2.0.0
Pillow==9.2.0
torch==2.0.0
torchvision==0.15.1
tqdm==4.65.0

@incollection{guo2024artwork,
  title={Artwork Protection Against Neural Style Transfer Using Locally Adaptive Adversarial Color Attack},
  author={Guo, Zhongliang and Dong, Junhao and Qian, Yifei and Wang, Kaixuan and Li, Weiye and Guo, Ziheng and Wang, Yuheng and Li, Yanli and Arandjelovi{\'c}, Ognjen and Fang, Lei},
  booktitle={ECAI 2024},
  year={2024},
  publisher={IOS Press}
}