Internet Traffic Analysis in a Large University Town: A Graphical and Clustering Approach

Abstract

Campus networks consist of a rich diversity of end hosts including wired desktops, servers, and wireless BYOD devices such as laptops and smartphones, which are often compromised in insecure networks. Making sense of traffic behaviors of end hosts in campus networks is a daunting task due to the open nature of the network, heterogeneous devices, high mobility of end users, and a wide range of applications. To address these challenges, this paper applies a combination of graphical approaches and spectral clustering to group the Internet traffic of campus networks into distinctive traffic clusters in a divide-and-conquer manner. Specifically, we first model the data communication between a particular subnet of campus networks and the Internet on a specific application port via bipartite graphs, and subsequently use the one-mode projection to capture behavior similarity of end hosts in the same subnet for the same network applications. Finally we apply a spectral clustering algorithm to explore the behavior similarity to identify distinctive application clusters within each subnet. Our experimental results have demonstrated the benefits of our proposed method for analyzing Internet traffic of a large university town to discover anomalous behaviors and to uncover distinctive temporal and spatial traffic patterns.