AnEvaluation ofNafve Bayesian Anti-Spai Filtering Techniques
Vikas P. Deshpande, Robert F. Erbacher · 2007
wordthat occurs onlyinspam,only inlegitimate messages, Abstract- Anefficient anti-spam fi'lter thatwouldblockall andinboth. Basedonthese wordoccurrence statistics (also spam,without blocking anylegitimate messages isa growingcalled tokens), incoming unseen messages areprocessed and need. Toaddress this problem, weexamine theeffectiveness of classified accordingly. statistically-based approaches NaiveBayesian anti-spam filters, asitiscontent-based andself-learning (adaptive) innature. Additionally, wedesigned aderivative filter basedonrelative II.BAYESIANCLASSIFIERS numbersoftokens. We train thefilters using alarge corpus of legitimate messages andspamandwetest thefilter using new A. Bayesian Classifier incoming personal messages. Morespecifically, fourfiltering A Bayesian classifier istheapplication ofa Bayesian techniques available foraNaive Bayesian filter areevaluated, network totheprocess oftextclassification. Bayesian We lookattheeffectiveness ofthetechnique, andweevaluatenetworks areprobabilistic networks that areusedasproblem different threshold values inordertofind anoptimal anti-spami filter configuration. Basedon cost-sensitive measures, we In odels idifferen felwork. conclude thatadditional safety precautions areneededfora Inourcase,a Bayesian networkisusedtorepresent a Bayesian anti-spam filter tobeputintopractice. However, our probability distribution ofspecified textcontained ina spam technique canmakeapositive contribution asafirst pass filter.email. Insuchagraph, anoderepresents arandomvariable, andadirected edgeindicates aprobabilistic dependency from IndexTerms-Spam filter, Naive Bayesian, Evaluation thevariable denoted bytheparent nodetothat ofthechild. Hence,itisimplied thatanynodeinthenetwork is conditionally independent ofitsnon-descendents, given its S PAMcontinues tobeagrowing problem accounting for parents. Eachnodeisassociated withaconditional probability upwards of90%ofalle-mail today (15). Whilespam table thatindicates thedistribution overthat nodewithany filters havebecomemoreeffective (16)andwidespread, possible assignment ofvalues toits parents (10, 13). manyspammessages continue tobedelivered toendusers. We formulate theBayesian networkto solveour Thedifficulty ineliminating spamlies indifferentiating it classification problem. LetC betheclass variable that fromalegitimate message. However, themessage content of indicates towhichclass (legitimate / spam)a message spamtypically formsadistinct category rarely observed in belongs, andletnodeXidenote anyattribute (token, inour legitimate messages, making itpossible fortextclassifiers to case) inthemessage. Forourpurposes wewill sayckisthe beusedforanti-spam filtering. Thegoal ofthis research isto givenofthespecific values fortherequired attributes. The examine theeffectiveness ofNaiveBayesian anti-spam filters specific values wouldbe0or1depending ontheir presence in andtheeffect ofparameter settings ontheeffectiveness of themessage. Theproblem ofclass nature canbesolved using spamfiltering. Additionally, we lookatanovel modification Baye's theorem: toexisting filters andincorporate itinto theevaluation. P(C kX-) P(X=xC=ck)P(C=ck)