search

GuanlinLee / warfare

Code for paper: Warfare:Breaking the Watermark Protection of AI-Generated Content (https://arxiv.org/abs/2310.07726)
2 stars 0 forks source link

Watermark extraction rate does not decrease #1

Closed zrsandwj1314 closed 3 months ago

zrsandwj1314 commented 3 months ago

image image Hello, according to the experimental setup in your paper on the Celeba dataset, the StegaStamp training effect, diffusion model output under large noise, and watermark extraction effect are basically the same. However, the watermark extraction rate has not decreased during the training of the generative adversarial network. I see that your paper showed a decrease at 400 epochs.

GuanlinLee commented 3 months ago

It seems the Dloss is too low. You can try to increase the weights of the discriminator. We notice the training process of the GAN is not stable. So, try more weight combinations, which are related to your devices and datasets. The hyperparameters in our paper are used in our experiments, but could not be suitable for all users.

zrsandwj1314 commented 3 months ago

Thank you very much for your answer. I will try more weight combinations. Thank you again for your research. I have learned a lot!

GuanlinLee commented 3 months ago

Thanks for your interest.