Adversarial Game Theory

"Security begins when the opponent is modeled as adaptive, not random."

Overview

Adversarial game theory studies attacker-defender interaction, robust objectives, strategic adaptation, and security games for AI systems.

Game theory is the part of the curriculum that studies adaptive decision makers. It asks what happens when each model, user, attacker, defender, or agent optimizes while anticipating the choices of others.

This section is written in LaTeX Markdown. Inline mathematics uses $...$, and display equations use `

`. The notes emphasize strategy, payoff, best response, equilibrium, exploitability, and adversarial adaptation.

Prerequisites

• Minimax Theorem
• Robustness and Distribution Shift
• Policy and Guardrails
• Model Serving and Inference Optimization

Companion Notebooks

Learning Objectives

After completing this section, you will be able to:

• Define attacker-defender games with actions, utilities, losses, and threat sets
• Write robust-risk objectives as nested minimization and maximization problems
• Explain perturbation sets for adversarial examples and PGD-style inner maximization
• Compute simple attacker best responses against defensive allocations
• Distinguish simultaneous, Stackelberg, and repeated security-game timing
• Use randomization and deception as strategic defensive tools
• Model GANs, red-team loops, benchmark gaming, and reward hacking as adaptive games
• Connect adversarial training to robustness under specified threat models
• Explain model extraction, poisoning, and jailbreak pressure as strategic adaptation
• Design evaluation protocols that account for adaptive opponents

Study Flow

1. Read the pages in order and pause after each page to restate the main definition or theorem.
2. Run theory.ipynb when you want to check the formulas numerically.
3. Use exercises.ipynb after the reading path, not before it.
4. Return to this overview page when you need the chapter-level navigation.

Runnable Companions

• Theory notebook
• Exercises notebook