Discovering and analyzing deviant communities: Methods and experiments

Abstract

Botnets continue to threaten the security landscape of computer networks worldwide. This is due in part to the time lag present between discovery of botnet traffic and identification of actionable intelligence derived from the traffic analysis. In this article we present a novel method to fill such a gap by segmenting botnet traffic into communities and identifying the category of each community member. This information can be used to identify attack members (bot nodes), command and control members (Command and Control nodes), botnet controller members (botmaster nodes), and victim members (victim nodes). All of which can be used immediately in forensics or in defense of future attacks. The true novelty of our approach is the segmentation of the malicious network data into relational communities and not just spatially based clusters. The relational nature of the communities allows us to discover the community roles without a deep analysis of the entire network. We discuss the feasibility and practicality of our method through experiments with real-world botnet traffic. Our experimental results show a high detection rate with a low false positive rate, which gives encouragement that our approach can be a valuable addition to a defense in depth strategy.