A1267
Title: Uncertainty quantification in latent position graph models
Authors: Nick Heard - Imperial College London (United Kingdom) [presenting]
Abstract: From a graph-based perspective, anomaly detection techniques currently deployed in enterprise cyber-security typically act on individual nodes or edges, for example, tracking connectivity patterns of a network host over time or detecting unusual volumes or periodicity in data transfers between two network nodes. Techniques which leverage the full network graph are less common; global network models have typically proved too simplistic in their assumptions, such as the well-studied but arguably overused stochastic block model. A new anomaly detection framework is proposed, which seeks to fully quantify uncertainty in node positions for latent position network graph models. Such a framework admits the possibility for nodes to be identified as outlying through, for example, unusual entropy levels in their perceived graph position rather than simply relying on detecting spatial outliers.
