Long before chatbots, machine learning models in fraud detection, malware classification and computer vision were being fooled, poisoned and copied. Those attacks still apply, now joined by their generative AI counterparts. This course teaches adversarial machine learning with the vocabulary of NIST AI 100-2 E2025 (attacker goals, knowledge and capabilities across the lifecycle) and maps it to MITRE ATLAS techniques.
You will implement gradient-based evasion (FGSM, PGD, Carlini-Wagner), black-box and transfer attacks, and domain-specific evasion against tabular fraud models and malware classifiers where feature constraints matter. You will study availability and targeted poisoning, backdoor trigger design, and poisoning of web-scale, fine-tuning and instruction data. The extraction module covers stealing models through prediction APIs and distilling LLM behavior, with defenses such as rate limiting, output perturbation and watermarking. The defense module treats adversarial training, certified robustness and randomized smoothing, and above all how to evaluate defenses correctly against adaptive attacks and gradient masking, using the Adversarial Robustness Toolbox and RobustBench. The course closes with prioritizing robustness work, building it into model release, and a capstone on a hypothetical fraud classifier. Privacy attacks are covered in sec-330 and application-level prompt attacks in sec-210. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by NIST, MITRE, the maintainers of the Adversarial Robustness Toolbox or RobustBench, or any vendor named here.