Adversarial machine learning examines how models fail when an attacker manipulates inputs, training data, queries, or surrounding tools. This course develops a broad security perspective across evasion, poisoning, privacy attacks, language-model threats, and agent risks, while connecting attacks to defenses and rigorous evaluation. It is suited to practitioners who need to reason about robustness throughout the machine learning lifecycle, from threat modeling and metrics to deployment monitoring and automated testing.
The course covers threat modeling, robustness metrics, white-box and black-box evasion, adversarial training, certified defenses, evaluation methods, poisoning and backdoors, privacy attacks, and supply-chain risks. It also covers jailbreaks, prompt injection, agent and MCP security, multimodal attack surfaces, LLM defenses, and automated red teaming.