Duel-based neuroevolutionary method for Stackelberg Security Games with boundedly rational Attacker

Abstract

This paper considers the problem of finding optimal strategies in Stackelberg Security Games when playing against a non-perfectly rational Attacker. To this end, a novel Duel-based NeuroEvolutionary approach to Security Games (DNESG) is proposed, which utilizes the Strategy Comparison Neural Network (SCNN) as a surrogate model to compare pairs of Defender’s strategies. SCNN is trained on historical data (past attack attempts) and does not require any direct information about the Attacker’s preferences regarding targets, payoff distribution, or decision-making model. SCNN is embedded in the Evolutionary Algorithm framework and implements a tournament-based selection method in place of a time-consuming direct strategy evaluation. The effectiveness of DNESG is assessed on a set of 90 benchmark Deep Packet Inspection games inspired by real cybersecurity scenarios. The proposed method provides high-quality solutions and outperforms state-of-the-art approaches (both exact and approximate) with statistical significance when playing against non-perfectly rational Attacker. Moreover, DNESG offers excellent time scalability, being two orders of magnitude faster than the state-of-the-art Mixed-Integer Linear Programming method.