ViTViz: Vision Transformer Adversarial Attack & Attention Visualization Tool

Execute adversarial attacks and visualize how attention evolves

Upload Image

Or choose a sample image:

Sample Image Carousel (click to select)

Attack Configuration

Attack Type

PGD/MIM/TGR: iterative | FGSM: single-step | SAGA variants: gradient × attention

Epsilon (ε) - Maximum Perturbation

0 1

Alpha (α) - Step Size

0 0.1

Steps - Number of Iterations

1 100

Metrics in Report

Choose which perturbation metrics should be displayed. Select All enables all metrics.

L∞ Distance

Attention W1 (Rollout)

PSNR

SSIM

LPIPS

Modified Pixels

Confidence Drop

Top-K Accuracy Drop

Attack Result

🎯

No attack executed yet

Configure the attack parameters above and click "Execute Full Analysis" to start. Results will appear here including prediction comparison, perturbation metrics, and attack success status.

Attention Evolution During the Attack

Compare the original image (left) with the attack iterations (right)

Visualization Controls

Mode

Alpha Overlay

0 1

Layer Index

Usado no modo Head

0 11

Head Index

Usado no modo Head

0 15

Select Attack Iteration to View

Attack Iteration

0 for the original image, last is the final adversarial image

0 10

Show Attention Heatmap

Overlay attention map on images

Original Image

Adversarial Image (Iteration)