Exploiting redundancy in natural language to penetrate Bayesian spam filters

摘要: Today's attacks against Bayesian spam filters attempt to keep the content of mails visible humans, but obscured filters. A common technique is fool by appending additional words a mail. Because these appear very rarely in mails, are inclined classify mail as legitimate. The idea we present this paper leverages fact that natural language typically contains synonyms. Synonyms different describe similar terms and concepts. Such often have significantly probabilities. Thus, an attacker might be able penetrate replacing suspicious innocuous with same meaning. precondition for success such attack users assign probabilities tokens. We first examine whether met; afterwards, measure effectivity automated substitution creating test set messages tested SpamAssassin, DSPAM, Gmail.