Latent representation
The input is encoded with the VAE encoder and represented in latent space.
AI IMAGE PROTECTION · RESEARCH PROTOTYPE
Protect images from AI reconstruction.
NoiseGuard explores adversarial image perturbations that aim to alter learned representations while keeping the visible image close to the original.
Stable Diffusion VAE encoder/decoder · PGD · MIRAGE surrogate
THE IDEA
Change the representation without visibly changing the image.
The prototype uses a frozen VAE to work with image representations, then applies two optimization stages before decoding the protected image.
01Input image
↓
02VAE encoding
↓
03PGD optimization
↓
04MIRAGE optimization
↓
05VAE reconstruction
HOW IT WORKS
A small research pipeline with measurable outputs.
Bounded optimization
PGD and MIRAGE optimize the image while keeping perturbations within the configured budget.
Verification
The tool returns the protected image, reconstruction, latent statistics, and optimization loss information.
READY TO TEST