A modular approach towards image spam filtering using multiple classifiers

Image based spam is a recent trick developed by the spammers´ community with the intention of bypassing the successful text based spam filters. Most of the traditional text based filters have been based on Naïve Bayes classification combined with text categorization methods. This work concentrates in developing a spam filtering system that accurately blocks image spam. The system analyzes images sent as attachments extracting both textual and visual features. The rationale behind employing a combination of both kinds of features is that spammers usually embed the payload in an image hidden by various obscuring methods. We used SVM classifier for the classification of low level features. The use of a noncommercial OCR for extracting text from images also delivered better accuracy. The Voting scheme provides a final measure of the spamminess of the images with its decision based on the maximum probability assigned by the two classifiers.