The code for ECCV2022 (Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal)
- The code is the implementation of the paper Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal.
- This paper has been received by ECCV 2020.
To install requirements:
- We use the CLWD (Colored Large-scale Watermark Dataset) in our experiments, which contains three parts: watermark-free images, watermarks and watermarked images. We first pretrain the watermark-removal networks using watermarked images in the train set of CLWD. Then in the attack stage, we use the watermark-free images as host images to generate watermark vaccines, and then add the watermarks with generated watermark vaccines.
- You can download pretrained models here: WDNet
python demo.py --model WDNet --epsilon 8 --start_epsilon 8 --step_alpha 2 --seed 160 --num_img 20 --attack_iter 50
- The first column is the input, the second is the output, and the last column is the predicted mask.
- First row is the clean image as an input, second row is the random noise input, last two rows are DWV and IWV respectively.
This work builds on many excellent works, which include: