Adversarial Examples for Preventing Diffusion Models from Malicious Image Edition

In recent years, with the advancement of artificial intelligence technology, Diffusion Models have become a prominent research direction, exhibiting remarkable proficiency in image generation tasks. However, the unrestricted utilization of Diffusion Models by infringers to illicitly edit unauthorized images has given rise to novel copyright challenges and privacy apprehensions. To address these issues, this paper introduces an adversarial sample-based approach that can significantly mitigate malicious image modifications by Diffusion Models. The key idea is to add imperceptible adversarial perturbations on the image, so that the representation of the disturbed image in the latent space is far away from the original image, thus effectively disrupting the editing operations of Diffusion Models and generating unrealistic pictures. A substantial volume of experimental results demonstrate the efficacy and robustness of this method.