The code "checkpoint = torch.load('./saved_models/cifar10/ckpt_blended_over.pth')" in zeromark.py. Does “ckpt_blended_over.pth” refer to the surrogate model trained by the attacker using the protected (the defender has already watermarked) dataset? How to watermark the dataset? How to calculate the cosine similarity between the boundary gradient and the watermark pattern?
Looking forward to your reply. Thank you very much. @JunfengGo
The code "checkpoint = torch.load('./saved_models/cifar10/ckpt_blended_over.pth')" in zeromark.py. Does “ckpt_blended_over.pth” refer to the surrogate model trained by the attacker using the protected (the defender has already watermarked) dataset? How to watermark the dataset? How to calculate the cosine similarity between the boundary gradient and the watermark pattern?
Looking forward to your reply. Thank you very much. @JunfengGo