E-Mail Spam Filter Based on Unsupervised Neural Architectures and Thematic Categories: Design and Analysis

Spam, or unsolicited messages sent massively, is one of the threats that affects email and other media. Its huge quantity generates considerable economic and time losses. A solution to this issue is presented: a hybrid anti-spam filter based on unsupervised Artificial Neural Networks (ANNs). It consists of two steps, preprocessing and processing, both based on different computation models: programmed and neural (using Kohonen SOM). This system has been optimized by utilizing a dataset built with ham from “Enron Email” and spam from two different sources: traditional (user’s inbox) and spamtrap-honeypot. The preprocessing was based on 13 thematic categories found in spams and hams, Term Frequency (TF) and three versions of Inverse Category Frequency (ICF). 1260 system configurations were analyzed with the most used performance measures, achieving \(\mathrm{AUC}>0.95\) the optimal ones. Results were similar to other researchers’ over the same corpus, although they utilize different Machine Learning (ML) methods and a number of attributes several orders of magnitude greater. The system was further tested with different datasets, characterized by heterogeneous origins, dates, users and types, including samples of image spam. In these new tests the filter obtained \(0.75<\mathrm{AUC}<0.96\). Degradation of the system performance can be explained by the differences in the characteristics of the datasets, particularly dates. This phenomenon is called “topic drift” and it commonly affects all classifiers and, to a larger extent, those that use offline learning, as is the case, especially in adversarial ML problems such as spam filtering.