Bayes Evasion Theory

However, there is an attack that does work

Choose words that more closely match conversational English

Ensure probabilities of a given word appearing are close to those in "nonspam" text

Force users to learn from those spams

Biases nonspam results towards spam

Tech Report at http://snipurl.com/97qb

(if that snipurl expires, Google for: "A Linguistics-Based Attack on Personalised Statistical E-mail Classifiers", Henry Stern, Justin Mason, and Michael Shepherd, Technical Report CS-2004-06, Dalhousie University, March 25, 2004)