adversarial-robustness-toolbox

Why we included this project

Teams responsible for the security of deployed machine learning models will find ART one of the most complete toolkits for testing how those models hold up under attack. It bundles a large catalog of adversarial techniques covering evasion, poisoning, extraction, and inference, along with matching defenses, detectors, and metrics, so you can run an attacker's playbook against your own model before someone else does. Because it spans the major frameworks and data types under one API, the same workflow applies whether you are testing a tabular classifier or an image object detector. The pairing of attacks with countermeasures is what makes it practical: generate a threat, run the defense, and get a quantified before-and-after in a single pass. Teams that want concrete robustness numbers instead of ad-hoc scripting get a ready-made reference library.