Several research projects try to protect images from AI. They share one idea - adversarial perturbations - but aim at different problems.
The shared idea: adversarial perturbations
AI models turn pixels into internal features. Small, carefully designed pixel changes - invisible to people - can push those features somewhere else entirely, so the AI "sees" a different picture.
PhotoGuard (MIT, 2023)
Targets AI editing with diffusion models. It perturbs a photo so the model's image encoder produces a wrong internal representation, which makes AI edits of that photo come out unrealistic. The published code needs a GPU and a large model.
Glaze (University of Chicago)
Protects artists' styles from being copied when their work is used to fine-tune image generators. It changes how a model perceives the artwork's style.
Nightshade (University of Chicago)
A "poisoning" approach aimed at AI training: images teach models wrong associations if they are scraped into training data without permission.
Where Photo Shield fits
Photo Shield is a free, lightweight tool for everyday photos. It runs the PhotoGuard encoder attack against TAESD, a small copy of the Stable Diffusion image encoder, so it works in your browser without uploading your photo. It is weaker than the full GPU research version and is best seen as an easy extra layer of protection, not a guarantee.
Limits all of these share
- Protection can weaken with compression, resizing or new AI models.
- None makes an image impossible to misuse.
- They work best combined with good privacy habits.
🛡️ Protect a photo now - free, in your browser
Last updated: October 2026