Adversarial Attack Lab
Nudge an image by a few invisible pixels and watch an image classifier change its mind.
Coming soon
This tool is being built. It will run right here, in your browser.
How it works
Method
Gradient-based attacks compute the smallest change that flips the model's prediction, shown side by side with the original.
Data
Sample photos, plus any image you load; nothing leaves your browser.
Model
A pre-trained image classifier, attacked with FGSM and PGD.
Build
In-browser model inference with a slider for attack strength.