Keyhole Adversarial Plan Recognition for Recognition of Suspicious and Anomalous Behavior

Adversarial plan recognition is the use of plan recognition in settings where the observed agent is an adversary, having plans or goals that oppose those of the observer. It is one of the key application areas of plan recognition techniques. There are two approaches to adversarial plan recognition. The first is suspicious activity recognition, i.e., directly recognizing plans, activities and behaviors that are known to be suspect (e.g., carrying a suitcase, then leaving it behind in a crowded area). The second is anomalous activity recognition, in which we indirectly recognize suspect behavior, by first ruling out normal, non-suspect, behaviors as explanations for the observations. Different challenges are raised in pursuing these two approaches. In this work we discuss a set of efficient plan recognition algorithms that are capable of handling the variety of challenges required of realistic adversarial plan recognition tasks. We describe an efficient hybrid adversarial plan recognition system that is composed of two processes: a plan recognizer capable of efficiently detecting anomalous behavior, and a utility-based plan recognizer incorporating the observer’s own biases—in the form of a utility function—into the plan recognition process. This allows choosing recognition hypotheses based on their expected cost to the observer. These two components form a highly efficient adversarial plan recognizer capable of recognizing abnormal and potentially dangerous activities. We evaluate the system with extensive experiments, using real-world and simulated activity data, from a variety of sources.