Improving Image Spam Filtering Using Image Text Features.
Giorgio Fumera, Fabio Roli, Battista Biggio, Ignazio Pillai · Conference on Email and Anti-Spam · 2008
Image-based spam (shortly, image spam) is a trick introduced by spammers few years ago. It consists in embedding all the textual information (i.e., the spam message) into an image attached to the spam e-mail. This allows to evade any filtering module based on the analysis of text in the e-mail’s body (usually, a naive Bayes classifier or a keyword detector). OCR-based modules have been proposed against image spam [7], and simple implementations have been included in spam filters like the popular SpamAssassin. However, besides requiring a relatively high processing time, OCR-based approaches are effective only for clean images, as shown in [4] for SpamAssassin. For this reason, spammers often obfuscate the text embedded into images, making OCR-based approaches ineffective. Other authors proposed to exploit image classification techniques to discriminate between ham and spam images (namely, images attached to ham and spam e-mails), using low-level visual features related for instance to colour distribution, to characteristics of text regions inside an image and to image meta-data [9, 1, 6]. In principle, this approach could be unaffected by text obfuscation techniques used by spammers and has a lower computational cost than OCR-based approaches (especially if classification algorithms as decision trees are used). Classification accuracies between 0.8 and 0.9 were reported in [1, 6] on real and artificial data sets of ham and spam images.