PhotoGuard AI
MIT research write-up and demo for PhotoGuard, using adversarial perturbations to immunize images against AI editing.
What it does
This is an MIT (gradient science) research post and demo for PhotoGuard, a technique that adds imperceptible adversarial perturbations to an image so diffusion models like Stable Diffusion cannot realistically edit it. It presents two approaches, attacking the model's conditioning step and the full diffusion process, along with a paper, code and interactive demo.
Core features
Adversarial immunization of images against AI editing
Two methods: conditioning attack and end-to-end diffusion attack
Interactive demo
Open paper and code
Best for
→Protecting photos from malicious AI manipulation
→Researching adversarial defenses for diffusion models
→Demonstrating image immunization techniques