photoguard

Why we included this project

PhotoGuard is a research defense against a specific problem: someone taking an image you posted and running it through a diffusion model to produce a convincing fake. The idea is to add a small adversarial perturbation to the image before it is shared, so that when a tool like Stable Diffusion tries to edit it, the output comes out visibly distorted or unrelated to the original. The repo covers two attack strategies, a lightweight one on the image encoder and a stronger end-to-end diffusion attack, with runnable notebooks for image-to-image and inpainting pipelines. It is research code tied to a paper, so expect to work through notebooks and tune parameters rather than drop in a turnkey service, though the interactive Gradio demo makes it easy to see the effect before you commit to integrating it.