Improving Spam Detection Based on Structural Similarity

We propose a new spam detection algorithm that uses structural relationships between senders and recipients of email as the basis for spam detection. A unifying representation of users and receivers in the vectorial space of their contacts is constructed, that leads to a natural definition of similarity between them. This similarity is then used to group email senders and recipients into clusters. Historical information about the messages sent and received by the clusters is obtained by forwarding messages to an auxiliary spam detection algorithm and this information is used to reclassify messages. In the framework proposed, our algorithm aims at correcting misclassifications from an auxiliary algorithm. A simulation is performed based on actual data collected from an SMTP server from a large University. We show that our approach is able reduce false positives, produced by the auxiliary classification algorithm, up to about 60%.