Abstract

We study the problem of identifying botnets and the IP addresses which they comprise, based on the observation of a fraction of the global email spam traffic. Observed mailing campaigns constitute evidence for joint botnet membership, they are represented by cliques in the graph of all messages. No evidence against an association of nodes is ever available. We reduce the problem of identifying botnets to a problem of finding a minimal clustering of the graph of messages. We directly model the distribution of clusterings given the input graph; this avoids potential errors caused by distributional assumptions of a generative model. We report on a case study in which we evaluate the model by its ability to predict the spam campaign that a given IP address is going to participate in.