foolbox

home: true heroImage: /logo.png heroText: Foolbox tagline: "Foolbox: Fast adversarial attacks to benchmark the robustness of machine learning models in PyTorch, TensorFlow, and JAX" actionText: Get Started → actionLink: /guide/ features:

• title: Native Performance details: Foolbox 3 is built on top of EagerPy and runs natively in PyTorch, TensorFlow, and JAX.
• title: State-of-the-art attacks details: Foolbox provides a large collection of state-of-the-art gradient-based and decision-based adversarial attacks.
• title: Type Checking details: Catch bugs before running your code thanks to extensive type annotations in Foolbox. footer: Copyright © 2022 Jonas Rauber, Roland S. Zimmermann

What is Foolbox?

Foolbox is a Python library that lets you easily run adversarial attacks against machine learning models like deep neural networks. It is built on top of EagerPy and works natively with models in PyTorch, TensorFlow, and JAX.

import foolbox as fb

model = ...
fmodel = fb.PyTorchModel(model)

attack = fb.attacks.LinfPGD()
epsilons = [0.0, 0.001, 0.01, 0.03, 0.1, 0.3, 0.5, 1.0]
advs, _, success = attack(fmodel, images, labels, epsilons=epsilons)