Toolkit chevron_right Adversarial Sandbox

Adversarial Robustness Sandbox

Experiment with real-time adversarial attacks on MobileNet V2 (TensorFlow.js).

dark_mode

Target Image

upload Upload Custom

Attack Configuration

Attack Method

Fast Gradient Sign Method: Adds noise in the direction of the loss gradient.

Perturbation (ε / 255) 0.05

Subtle Obvious

info How it works

This tool runs MobileNet V2 in your browser. When you click "Execute Attack", we calculate the gradients of the model's loss with respect to the input pixels. We then add a tiny amount of noise (scaled by ε) to maximize the error, effectively pushing the image across the decision boundary.

check_circle Original Input

dangerous Adversarial Example

Run attack to see results

Noise Pattern (Amplified)

L₂ Distance: 0.0000

L∞ Distance: 0.0000

Noise is shifted to gray (127) + diff.

Confidence Shift

(Chart removed for performance)