Rethinking the Invisible Protection against Unauthorized Image Usage in Stable Diffusion

Advancements in generative AI models like Stable Diffusion, DALL·E 2, and Midjourney have revolutionized digital creativity, enabling the generation of authentic-looking images from text and altering existing images with ease. Yet, their capacity poses significant ethical challenges, including replicating an artist's style without consent, the creation of counterfeit images, and potential reputational damage through manipulated content. Protection techniques have emerged to combat misuse by injecting imperceptible noises into images. This paper introduces Insight, a novel approach that challenges the robustness of these protections by aligning protected image features with human visual perception. By using a photo as a reference, approximating the human eye's perspective, Insight effectively neutralizes protective perturbations, enabling the generative model to recapture authentic features. Our extensive evaluation across 3 datasets and 10 protection techniques demonstrates its superiority over existing methods in overcoming protective measures, emphasizing the need for stronger safeguards in digital content generation.