Adversarial Techniques

Image watermarks embed hidden messages that are only detectable by the owners, preventing misuse of images, especially those generated by AI models. In this paper we introduce a family of regeneration attacks to remove these invisible watermarks. By adding random noise to an image and then reconstructing it, this approach effectively eliminates watermarks. The study highlights the vulnerability of all invisible watermarks to this proposed attack, emphasizing the need for a shift toward semantically similar watermarking methods.