adversarial-attacks-pytorch

Why we included this project

Security engineers and ML teams who need to know whether a model breaks under hostile input will get more practical use out of this than a research paper's reference implementation. It wraps the classic attack algorithms, FGSM, PGD, CW, BIM and others, in a single PyTorch-style API, so switching perturbation methods means changing one class instantiation instead of rewriting the attack loop. That is handy when you are stress-testing a deployed classifier or checking how a model behaves under a specific attack budget. Attacks can run targeted or untargeted, multiple methods can be combined into an ensemble, and you can force the model into training or eval mode for edge cases like RNNs. The ability to save and reload generated adversarial examples also makes it easy to keep reproducible test sets for robustness reporting.