Adversarial perturbations cannot reliably protect artists from generative AI

The paper evaluates popular style imitation protection tools with millions of downloads and argues that they provide only a false sense of security. Low-cost off-the-shelf techniques such as image upscaling are sufficient to build robust imitation methods that significantly weaken existing protections. Through a user study, the authors show that all existing protections can be easily bypassed and that artists still face the risk of their style being imitated. The authors accordingly caution that tools based on adversarial perturbations cannot reliably protect artists from misuse of generative AI, and call for exploring non-technical alternatives.