How Spammers Fool Bayesian Filters - And How to Stop Them

by: Paul Judge, CTO, CipherTrust, Inc.
Effectively stopping spam over the long-term requires much more than blocking individual IP addresses and creating rules based on keywords that spammers typically use. The increasing sophistication of spam tools coupled with the increasing number of spammers in the wild has created a hyper-evolution in the variety and volume of spam. The old ways of blocking the bad guys just don’t work anymore. Examining spam and spam-blocking technology can illuminate how this evolution is taking place and what can be done to combat spam and reclaim e-mail as the efficient, effective communication tool it was intended to be. One method used to combat spam is Bayesian Filtering. Named after Thomas Bayes, an English mathematician, Bayesian Logic is used in decision making and inferential statistics. Bayesian Filers maintain a database of known spam and ham, or legitimate email. Once the database is large enough, the system ranks the words according to the probability they will appear in a spam message. Words more likely to appear in spam are given a high score (between 51 and 100), and words likely to appear in legitimate email are given a low score (between 1 and 50). For example, the words “free” and “sex” generally have values between 95 and 98, whereas the words “emphasis” or “disadvantage” may have a score between 1 and 4. Commonly used words such as “the” and “that”, and words new to the Bayesian filters are given a neutral score between 40 and 50 and would not be used in the system’s algorithm. When the system receives an email, it breaks the message down into tokens, or words with values assigned to them. The system utilizes the tokens with scores on the high and low end of the range and develops a score for the email as a whole. If the email has more spam tokens than ham tokens, the email will have a high spam score. The email administrator determines a threshold score the system uses to allow email to pass through to users. Bayesian filters are effective at filtering spam and minimizing false positives. Because they adapt and learn based on user feedback, Bayesian Filers produce better results as they are used within an organization over time. They are not, however, foolproof. Spammers have learned which words Bayesian Filters consider spammy and have developed ways to insert non-spammy words into emails to lower the message’s overall spam score. By adding in paragraphs of text from novels or news stories, spammers can dilute the effects of high-ranking words. Text insertion has also caused normally legitimate words that are found in novels or news stories to have an inflated spam score. This may potentially render Bayesian filters less effective over time. Another approach spammers use to fool Bayesian filters is to create less spammy emails. For example, a spammer may send an email containing only the phrase, “Here’s the link…”. This approach can neutralize the spam score and entice users to click on a link to a Web site containing the spammer’s message. To block this type of spam, the filter would have to be designed to follow the link and scan the content of the Web site users are asked to visit. This type of filtering is not currently employed by Bayesian filters because it would be prohibitively expensive in terms of server resources and could potentially be used as a method of launching denial of service attacks against commercial servers. As with all single-method spam filtering methodologies, Bayesian filters are effective against certain techniques spammers use to fool spam filters, but are not a magic bullet to solving the spam problem. Bayesian filters are most effective when combined with other methods of spam detection. The Solution When used individually, each anti-spam technique has been systematically overcome by spammers. Grandiose plans to rid the world of spam, such as charging a penny for each e-mail received or forcing servers to solve mathematical problems before delivering e-mail, have been proposed with few results. These schemes are not realistic and would require a large percentage of the population to adopt the same anti-spam method in order to be effective. You can learn more about the fight against spam by visiting our website at www.ciphertrust.com and downloading our whitepapers.