Good Word Attacks on Statistical Spam Filters

Unsolicited commercial email is a significant problem for users and providers of email services. While statistical spam filters have proven useful, senders of spam are learning to bypass these filters by systematically modifying their email messages. In a good word attack, one of the most common techniques, a spammer modifies a spam message by inserting or appending words indicative of legitimate email. In this paper, we describe and evaluate the effectiveness of active and passive good word attacks against two types of statistical spam filters: naive Bayes and maximum entropy filters. We find that in passive attacks without any filter feedback, an attacker can get 50% of currently blocked spam past either filter by adding 150 words or fewer. In active attacks allowing test queries to the target filter, 30 words will get half of blocked spam past either filter.