A machine learning approach to server-side anti-spam e-mail filtering

Abstract

Spam-detection systems based on traditional methods have several obvious disadvantages like low detection rate, necessity of regular knowledge bases’ updates, impersonal filtering rules. New intelligent methods for spam detection, which use statistical and machine learning algorithms, solve these problems successfully. But these methods are not widespread in spam filtering for enterprise-level mail servers, because of their high resources consumption and insufficient accuracy regarding false-positive errors. The developed solution offers precise and fast algorithm. Its classification quality is better than the quality of Naïve-Bayes method that is the most widespread machine learning method now. The problem of time efficiency that is typical for all learning based methods for spam filtering is solved using multi-agent architecture. It allows easy system scaling and building unified corporate spam detection system based on heterogeneous enterprise mail systems. Pilot program implementation and its experimental evaluation for standard data sets and for real mail flows have demonstrated that our approach outperforms existing learning and traditional spam filtering methods. That allows considering it as a promising platform for constructing enterprise spam filtering systems.