ABSTRACT Robot swarms can be deployed as moving surveillance systems, for instance, as mobile anti‐poaching systems for monitoring wildlife and detecting poaching activities. Since poachers have an interest in evading detection, robots are at risk of being hijacked and manipulated to behave antagonistically, for example, to prevent the correct surveillance of a group of animals in a certain area. While the detection of cyber‐attacks on robots is commonly considered in the context of network activity and communication, this work focuses on the detection of anomalies that manifest in the robots' physical motion behavior. It builds upon our previous work, where a contextual data‐driven anomaly detection method was proposed and applied to motions of a robot swarm monitoring a static area. Such data‐driven, unsupervised methods for detecting anomalies in the physical behavior of robots are a suitable way of evading the need to define behavioral signatures or rules but only use information on the behavior of “normal,” cooperating robots, and the method proposed earlier demonstrated a strong performance both for recognizing normal robots and for the detection of a single antagonistic robot within a swarm monitoring a static area. Motivated by anti‐poaching scenarios, this article goes beyond the earlier work by considering multiple antagonists and a dynamic scenario, where the robots monitor a continually changing area around a herd of animals and therefore do not converge to fixed, optimal positions as in a static scenario. The statistical evaluation shows that the anomaly detection method consistently yields a correct categorization of normal and antagonistic behavior, and that the method's performance is unaffected by an increased number of antagonistic robots in a swarm. Furthermore, the results indicate that, given the availability of a limited amount of new data, the method can be fine‐tuned to efficiently adapt to new settings.
Wenger et al. (Thu,) studied this question.