Game-Theoretic Models for Interdependent Security: Modeling, Computing, and Learning.

摘要

Due to an increase number of attacks by hackers and terrorists, there has been quite a bit of recent research activity in the general area of game-theoretic models for terrorism settings that aim to understand the behavior of the attackers and the attackers' targets. My thesis is centered on introducing, studying, and applying several game-theoretic models to security.;In particular, my doctoral thesis consists of the following components: (1) designing increasingly more realistic variants of defense games; (2) studying computational questions in defense games such as equilibria computation and computational implications of equilibria characterizations, (3) designing efficient algorithms and effective heuristics for defense problems; and (4) designing and applying new machine learning techniques to estimate game model parameters from behavioral data.;Our computational models build on top of interdependent security (IDS) games, a model introduced by economists and risk-assessment experts Kunreuther and Heal to study investment decisions of strategic agents when facing direct and transfer risk exposure from other agents in the system. We first introduce generalized IDS (alpha-IDS) games, a model that extends IDS games where full investment can reduce transfer risk. In particular, alpha is a vector of probabilities, one for each agent, specifying the probability that the transfer risk will not be protected by the investment. In other words, agent i's investment can reduce indirect risk by probability (1 -- alphai). We then extend from alpha-IDS games and introduce interdependent defense (IDD) games, a computational-game-theoretic framework for settings of interdependent security to study scenarios of multiple-defenders vs. a single-attacker in a network. For the variants of defense games we introduced, we study some computational aspects of computing Nash equilibria in those games.;Finally, we investigate the problem of learning the games form observed behavioral data. For this problem, we introduce a machine-learning generative model to learn the parameters of the games. As an application, we apply the learning model and use machine-learning techniques to estimate the parameters and structure of alpha-IDS games using the vaccination data from the Centers for Disease Control and Prevention (CDC) in the United States.