Flow-level Characteristics of Spam and Ham

Despite a large amount of effort devoted in the past years trying to limit unsolicited mail, spam is still a major global concern. Content-analysis techniques and blacklists, the most popular methods used to identify and block spam, are beginning to lose their edge in the battle. We argue here that one not only needs to look into the network-related characteristics of spam traffic, as has been recently suggested, but also to look deeper into the network core, in order to counter the increasing sophistication of spam-ing methods. Yet, at the same time, local knowledge available at a given server can often be irreplaceable in identifying specific spammers. To this end, in this paper we show how the local intelligence of mail servers can be gathered and correlated passively at the ISP-level providing valuable network-wide information. Specifically, we use first a large network flow trace from a medium size, national ISP, to demonstrate that the pre-filtering decisions of individual mail servers can be tracked and combined at the flow level. Then, we argue that such aggregated knowledge not only can allow ISPs to develop and evaluate powerful new methods for fighting spam, but also to monitor remotely what their “own” servers are doing.