Unit 4

Adversarial AI

AI systems can be accurate on ordinary test data and still fail on inputs chosen to make them fail. In this unit, you will attack small systems, trace how the attacks work, compare defenses, and build a practical red-team plan.

Chapter 1

Fooling an AI System

You will be able to recognize adversarial behavior, explain why ordinary accuracy misses it, build a threat model, and locate attack surfaces across an AI system.

Chapter 2

Adversarial Attacks

You will be able to classify attacks by lifecycle stage and attacker access, explain how FGSM and physical attacks work, trace a model-extraction pipeline, and distinguish poisoning, prompt injection, and sycophancy.

Chapter 3

Adversarial Defenses

You will be able to explain adversarial training, interpret a certified guarantee, describe how defensive distillation transfers robustness from teacher to student, and evaluate robustness against a clearly stated threat model.

Chapter 4

Red Teaming

You will be able to distinguish red and blue team responsibilities, design a scoped adversarial evaluation, and turn a model failure into reproducible evidence and a retestable mitigation.