Break neural networks in your browser.
An interactive, in-browser demonstration of adversarial attacks on neural networks – entirely in JavaScript.
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)
src/adversarial.js
: The core adversarial.js library that implements all attacks.docs/
: The interactive demo directory. Explore the live website.docs/js/
: Contains a ton of scripts that process data and power the demo, and a copy ofadversarial.js
.docs/data/
: Contains (subsets) of datasets used in the demo.