cleverhans

Why we included this project

CleverHans is a practical first stop for teams hardening machine-learning models against adversarial examples. It packages reference implementations of well-known attacks, so you can probe a model's weaknesses without reimplementing each method yourself. The library covers attacks and defenses in PyTorch, JAX, and TensorFlow 2, which helps when your stack spans more than one backend and you want results that carry across. Because it is built around benchmarking, you can test a model against standard attack methods and report figures others can reproduce. Security researchers and ML engineers checking a deployment against evasion-style attacks are the main audience, and the MNIST and CIFAR-10 tutorials with common attack methods keep the learning curve gentle. Treat it as a test harness for robustness checks rather than something to run in production.