A neural network classifier for junk e-mail

Most e-mail readers spend a non-trivial amount of time regularly deleting junk e-mail (spam) messages, even as an expanding volume of such e-mail occupies server storage space and consumes network bandwidth. An ongoing challenge, therefore, rests within the development and refinement of automatic classifiers that can distinguish legitimate e-mail from spam. A few published studies have examined spam detectors using Naive Bayesian approaches and large feature sets of binary attributes that determine the existence of common keywords in spam, and many commercial applications also use Naive Bayesian techniques. Spammers recognize these attempts to thwart their messages and have developed tactics to circumvent these filters, but these evasive tactics are themselves patterns that human readers can often identify quickly. Therefore, in contrast to earlier approaches, our feature set uses descriptive characteristics of words and messages similar to those that a human reader would use to identify spam. This preliminary study tests this alternative approach using a neural network (NN) classifier on a corpus of e-mail messages from one user. The results of this study are compared to previous spain detectors that have used Nave Bayesian classifiers. Also, it appears that commercial spam detectors are now beginning to use descriptive features as proposed here.