adversarial.js

Break neural networks in your browser. Just for the fun of it.

An interactive, in-browser demonstration of adversarial attacks on neural networks – entirely in JavaScript.

Get Started

Try the demo!

Implementation

adversarial.js is completely implemented in JavaScript – so it can run entirely within your browser. We rely on TensorFlow.js.

The library supports the following attacks:

• Fast Gradient Sign Method (L_inf attack)
• Basic Iterative Method (L_inf attack)
• Jacobian-based Saliency Map Attack (L0 attack)
• Jacobian-based Saliency Map Attack, One-Pixel (L0 attack)
• Carlini & Wagner (L2 attack)

The demo can break the following pre-loaded systems:

• MNIST (digit recognition)
• GTSRB (street sign recognition)
• CIFAR-10 (object recognition, small images)
• ImageNet (object recognition, large images)

FAQ

See the website.

Repo Structure

1. src/adversarial.js: The core adversarial.js library that implements all attacks.
2. docs/: The interactive demo directory. Explore the live website.
3. docs/js/: Contains a ton of scripts that process data and power the demo, and a copy of adversarial.js.
4. docs/data/: Contains (subsets) of datasets used in the demo.