Multiagent Communication Security in Non-Zero Sum Adversarial Settings

In future battlefields and other emerging multiagent applications, agents must communicate in an inherently hostile environment in which an adversary has strong incentives to disrupt or intercept communication. Intelligent agents must balance network performance with possible harm suffered from the adversary’s attacks, given that the adversary is actively and rationally balancing his own costs for attacking the network with the possible harm inflicted. The SAFER algorithm was proposed to meet these challenges in the absence of attack costs for the adversary, by representing the problem as a zero-sum game between a sender agent choosing communication paths through a network and an adversary choosing nodes to attack. However, this zero-sum assumption is often violated in real applications. In this paper we extend the previous approach to a non-zero sum game where attacking a node causes the adversary to incur a cost which factors into the adversary’s payoff but not into the sender’s payoff. We characterize properties of the equilibrium in this new game and provide a polynomial time algorithm for finding a Nash equilibrium. We also consider a sequential form of the game where the sender plays first and the adversary observes the sender’s action before launching his attack. We show that in this case the sender can actually improve his payoff, and provide a polynomial time algorithm for finding an equilibrium strategy for the sequential game when the adversary can attack a single node.