External links

Full text: HTML, PDF

Information about this paper

Abstract

Bayesian content filters gained popular acclaim when they were put forward in 2002 by Paul Graham as a potential long-term solution for the spam problem. They have since fallen from the limelight, however, due to perceived attack vulnerabilities inherent to all content-based filters as well as real and imagined vulnerabilities specific to Bayesian filters. It has also been assumed that Bayesian filters would be problematic to implement in centralized or large environments due to wordlist management issues. This paper revisits the effectiveness of Bayesian filters as a sustainable singular spam solution for mid- to large-sized environments through a real-world study of the deployment and operation of the Bogofilter Robinson-Fisher Bayesian classification utility in a production mail environment servicing thousands of accounts. Our implementation strategy and methodology as well as our results are described in detail so that they can be evaluated and replicated if desired. Other filtering methodologies which were previously implemented in this environment are also discussed for comparison purposes, though they have since been removed from production due primarily to lack of need. Bayesian classification has been able to solve the spam problem for this user population for the present and observable future, with a single wordlist, and with no secondary spam filtering techniques employed. Significantly, only two business-related legitimate messages have been reported as blocked due to filter misclassification since Bogofilter was deployed.