Robust classification for spam filtering by back-propagation neural networks using behavior-based features

Earlier works on detecting spam e-mails usually compare the contents of e-mails against specific keywords, which are not robust as the spammers frequently change the terms used in e-mails. We have presented in this paper a novel featuring method for spam filtering. Instead of classifying e-mails according to keywords, this study analyzes the spamming behaviors and extracts the representative ones as features for describing the characteristics of e-mails. An back-propagation neural network is designed and implemented, which builds classification model by considering the behavior-based features revealed from e-mails’ headers and syslogs. Since spamming behaviors are infrequently changed, compared with the change frequency of keywords used in spams, behavior-based features are more robust with respect to the change of time; so that the behavior-based filtering mechanism outperform keyword-based filtering. The experimental results indicate that our methods are more useful in distinguishing spam e-mails than that of keyword-based comparison.