Guarding Against Network Intrusions 99 words or word combinations from the message's header and body. Second, probabilities are assigned to tokens through a training process. The filter looks at a set of known spam messages compared to a set of known legitimate messages and calculates token probabilities based on Bayes' theorem (from probability theory). Intuitively, a word such as Viagra would appear more often in spam, and therefore the appearance of a Viagra token would increase the probability of that message being classified as spam. The probability calculated for a message is compared to a chosen threshold; if the probability is higher, the message is classified as spam. The threshold is chosen to balance the rates of false positives and false negatives (missed spam) in some desired way. An attractive feature of Bayesian filtering is that its probabilities will adapt to new spam tactics, given continual feedback, that is, correction of false positives and false negatives by the user. It is easy to see why spammers have attacked Bayesian filters by attempting to influence the probabilities of tokens. For example, spammers have tried filling messages with large amounts of legitimate text (e.g., drawn from classic literature) or random innocuous words. The presence of legitimate tokens tends to decrease a message's score because they are evidence counted toward the legitimacy of the message. Spammers are continually trying new ways to get through spam filters. At the same time,