Poisoning Attacks on Data-Driven Utility Learning in Games

Ruoxi Jia, Ioannis C. Konstantakopoulos, Bo Li, Costas J. Spanos

Abstract: Game theory has been employed for modeling agents' decisions in various science and engineering fields. Game theoretic analysis often assumes that the utility function of each agent is known a priori, and yet this assumption does not hold for many real-world applications. The combination of Internet of Things (IoT) and advanced data analysis techniques has stimulated fruitful research on learning agents' utility functions from data. Just as many other data-driven methods, utility learning also suffers from potential security risks. Due to the great economic value of accurate forecasting of agents' behaviors, there are huge incentives for adversaries to attack utility learning methods by poisoning training datasets and mislead predictions to achieve malicious goals. In this paper, we introduce and analyze optimal poisoning attack strategies in order to understand adversarial actions and further encourage potential defenses. Moreover, we study how an adversary might disguise the attacks by mimicking normal actions. The proposed attack strategies are evaluated on both synthetic and real-world social energy game data, and the results show that the root mean squared error in predicting agents' actions increases by up to 67 % by adding only 5 % well-crafted poisoning training instances.