bethgelab/foolbox

Health v2 · maintenance only

Adoption not part of the score

2972 stars · 442 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Foolbox is a Python library for generating adversarial examples that fool deep neural networks, with state-of-the-art gradient-based and decision-based attacks. Built on EagerPy, it runs natively on PyTorch, TensorFlow, and JAX models with a single codebase and batch support.

Use cases

• generate adversarial examples to fool a neural network
• benchmark robustness of my PyTorch model against attacks
• run PGD attacks on an ImageNet classifier
• test model robustness in TensorFlow or JAX
• evaluate robust accuracy across multiple epsilon values
• compare gradient-based and decision-based adversarial attacks

When to choose

• you need adversarial attacks that work natively across PyTorch, TensorFlow, and JAX
• you want a well-documented, actively maintained attack toolbox with batch support
• you need state-of-the-art gradient-based and decision-based attacks in one library

When to avoid

• you need adversarial training/defense methods rather than attacks
• you work with frameworks other than PyTorch, TensorFlow, or JAX
• you only need a simple FGSM one-liner and don't want a full toolbox

Facets

library · maturity active

machine-learning security benchmarking testing machine-learning deep-learning security artificial-intelligence python cross-platform adversarial-examples adversarial-attacks robustness-benchmarking pytorch tensorflow jax eagerpy deep-learning-security

Member repositories

For agents

markdown · JSON · MCP: product_card(name="bethgelab/foolbox")