COMBINED ATTACK
NoiseGuard Protection Pipeline
VAE → PGD → MIRAGE → VAE Decoder. The graphs below plot the actual values returned by the backend for every attack step.
01VAE Encode
→
02Latent PGD
→
03MIRAGE
→
04VAE Decode
1. Load and Run Combined Attack
Original
No image loaded
![Original image]()
VAE Reconstruction
Waiting for run
![VAE reconstruction]()
Protected
Waiting for run
![Protected image]()
2. Latent Statistics
-
Latent Mean
-
Latent Std
-
Latent Shape
-
Reconstruction MSE
-
Latent Distance
3. Combined Attack Curves
Each graph uses the raw loss history returned by the backend. The axes automatically scale to the observed values, so a small change is not hidden by a large empty range.
PGD latent-distance loss
MIRAGE target similarity
EVALUATION
4. Test Against an Open-Source Image Model
The backend sends the protected image to Stable Diffusion v1.5 Img2Img as a downstream reconstruction test. This is separate from the NoiseGuard attack itself.
BACKEND + OPEN-SOURCE MODEL TEST
Original
NoiseGuard Protected
Stable Diffusion Reconstruction
NoiseGuard metrics
-
Latent Distance
-
L∞ Perturbation
-
Budget Used
-
Protected PSNR
-
MIRAGE Score Change
Downstream model reconstruction
-
Model
-
Generated vs Original PSNR
-
Generated vs Original MSE
-
Generated vs Protected PSNR
-
Generated vs Protected MSE
Status: Idle — load an image to begin