foolbox

Why we included this project

Adversarial robustness is easy to ignore until a model is already in production and someone finds a way to break it. Foolbox is a Python library for probing that weakness before deployment: it runs a large collection of gradient-based white-box and decision-based black-box attacks against your own models. Since it's built on EagerPy, the same attack code works on PyTorch, TensorFlow, and JAX, so you can test robustness consistently even when your stack mixes frameworks. It reports adversarial examples and success rates across different perturbation budgets, which makes it straightforward to compare how architectures or defenses hold up. For ML engineers and researchers who care about model robustness, it's a well-documented, practical place to start.