Real-time detection of anomalous paths through networks

Abstract

The proliferation of increasingly inexpensive mobile devices capable of transmitting accurate positional information to other devices and servers has led to a variety of applications ranging from health situation monitoring to GPS-based offender monitoring. One of the resultant challenges is in understanding, in real-time, when incoming observations merit further examination. In this research, we investigate an approach for identifying anomalous paths through networks using real-time comparisons to a previously learned model. Our approach, the development of a series of “posterior weighted graphs” allows us to both determine which underlying model a particular path most closely represents as well as evaluate this relationship in real-time as more observations become available. Here we present the posterior weighted graph approach for examining path similarity and an extension for detecting anomalies in real-time. Our results illustrate how we can distinguish from among multiple candidate paths and, likewise, when observations no longer match an expected model.