Spam Filtering

Anirudh Ramachandran (Nick Feamster/Santosh Vempala) - Spam filtering
Our project, SpamTracker, is a system for early detection of spamming bots that capitalizes on spamming bots behaving similar to other bots, whereas legitimate email senders do not. We have identified two metrics of behavior for email senders: (1) Correlation of sending behavior across domains targeted by an email, and (2) Correlation of sending behavior across time. We constructed a classifier for metric #1 by encoding sender behavior across target domains as an object-feature matrix, and applying fast clustering to identify sending patterns that contained known spammers. We used these patterns (encoded as ``behavioral fingerprints'') to classify previously unknown senders that matched the patterns as spammers. We found that our classifier was able to automatically detect many spammers quicker than other common techniques that had access to human input. The project for the current semester intends to extend and improve this work by 1) making the classifier more accurate, 2) optimizing it to function at line rate.