Collusion Set Detection Through Outlier Discovery.

The ability to identify collusive malicious behavior is critical in today's security environment. We pose the general problem of Collusion Set Detection (CSD): identifying sets of behavior that together satisfy some notion of "interesting behavior". For this paper, we focus on a subset of the problem (called CSD′), by restricting our attention only to outliers. In the process of proposing the solution, we make the following novel research contributions: First, we propose a suitable distance metric, called the collusion distance metric, and formally prove that it indeed is a distance metric. We propose a collusion distance based outlier detection (CDB) algorithm that is capable of identifying the causal dimensions (n) responsible for the outlierness, and demonstrate that it improves both precision and recall, when compared to the Euclidean based outlier detection. Second, we propose a solution to the CSD′ problem, which relies on the semantic relationships among the causal dimensions. [ABSTRACT FROM AUTHOR]