Graduate student Anirudh Ramachandran's work on filtering spam using
network-level properties will appear at the ACM Conference on Computer
and Communications Security (CCS), ACM's top security conference, at
the end of October.  Ramachandran and his advisor, Assistant Professor
Nick Feamster, have been working with Professor Santosh Vempala to
develop next-generation spam filtering techniques.

Spam is becoming increasingly virulent as it makes use of images and
PDFs to evade content-based filters.  To make matters worse, spammers
are sending spam from "fresh" machines every day, which makes it
difficult to maintain static blacklists of known bad senders. 

To get a step ahead, the researchers have taken a different approach:

rather than filtering spam based on content or an ephemeral identity of
the sender (e.g., an IP address), the researchers have invented a new
technique called "behavioral blacklisting".  Behavioral blacklisting
aims to learn and "fingerprint" spammers' sending patterns---for
example, the set of recipients a particular sender is targeting---and
blacklist senders based on their sending behavior, rather than a fixed
identity.

The researchers developed their first behavioral blacklisting technique
by applying Professor Vempala's novel spectral clustering algorithms,
which have also successfully been applied to other areas (e.g., Web
search).

You can read the paper here.