Spectral analysis to anomaly detection in dynamic attributed graphs

Abstract

Anomaly detection in data has a wide variety of applications, such as intrusion detection in computer networks or credit card fraud detection. In particular, if the data can represented by graphs, metrics and methods based on graph theory known for high quality solutions can be used. In the case of static graphs, the anomaly detection problem has been extensively studied and several algorithms have been proposed. However, there are a few studies that aim at anomaly detection in dynamic attributed networks. In these cases, the majority of the existing strategies considers node and edge updates, ignoring the history of the changes along the anomaly detection process. In addition, few strategies are scalable to deal with Big Data. In this case, clustering-based anomaly detection strategies are pointed out as a good solution, since they allow to analyze groups of vertices instead of individual vertices and, therefore, have a lower computational cost. Thereby, this project aims atinvestigating the existing unsupervised methods for anomaly detection in dynamic attributed networks. As main contribution, a scalable strategy for anomaly detection in dynamic attributed networks will be proposed. This strategy will use a clustering algorithm and spectral operators that will also be developed during this project. The developed strategy will be tested to attest its efficiency and its results will be compared with those from the best strategies found in the literature. (AU)