Mitigating Deception and Interference in Online Goal Recognition Systems

Abstract

Online Goal Recognition (OGR) is the task of recognizing an agent's goal while that agent is executing a plan. Several OGR systems have been designed to observe and analyze an agent's actions in order to infer its goal. However, these systems generally assume that the agent is either unaware of being observed or is cooperating with the system. In this paper, we analyze two scenarios in which this assumption does not hold: Deception, where an agent deliberately hides its true goal by computing a deceptive plan; and Interference, where another agent tampers with the original plan. In particular, we evaluate the performance of the two state-of-the-art systems for OGR (ORL and CLERNet) in these two scenarios, gathering and extending different types of deceptive and interfering attacks. Moreover, we propose a framework (PAC-OGR) that mitigates the effect of the attacks by amending the manipulated plan and reasoning about the agent's behaviour. An experimental evaluation over several classical planning domains shows that PAC-OGR can be effectively integrated into existing OGR systems, making them more robust and reliable.