Security in network games

Attacks on the Internet are characterized by several alarming trends: 1) increases in frequency; 2) increases in speed; and 3) increases in severity. Modern computer worms simply propagate too quickly for human detection. Since attacks are now occurring at a speed which prevents direct human intervention, there is a need to develop automated defenses. Since the financial, social and political stakes are so high, we need defenses which are provably good against worst case attacks and are not too costly to deploy. In this dissertation we present two approaches to tackle these problems. For the first part of the dissertation we consider a game between an alert and a worm over a large network. We show, for this game, that it is possible to design an algorithm for the alerts that can prevent any worm from infecting more than a vanishingly small fraction of the nodes with high probability. Critical to our result is designing a communication network for spreading the alerts that has high expansion. The expansion of the network is related to the gap between the 1st and 2nd eigenvalues of the adjacency matrix. Intuitively high expansion ensures redundant connectivity. We also present results simulating our algorithm on networks of size up to $2^{25}$. In the second part of this dissertation we consider the virus inoculation game which models the selfish behavior of the nodes involved. We present a technique for this game which makes it possible to achieve the \u27windfall of malice\u27 even without the actual presence of malicious players. We also show the limitations of this technique for congestion games that are known to have a windfall of malice