Image spam is a kind of spam in which the text message is embedded into attached images to circumvent filtering modules based on the analysis of e-mails' body text. Approaches proposed so far against image spam rely on:

OCR-based filters: where text embedded into images is extracted through OCR tools and analyzed using standard techniques like simple keyword detection or statistical text categorization techniques. These approaches can be useful against those kinds of image spam in which images are clean (without obfuscation techniques like the ones used in CAPTCHAs).
Based on this approach, in collaboration with the P.R.A.Group Of the University of Cagliari, we have also developed a plugin for the open source filter SpamAssassin: BayesOCR plugin
The basic idea of ImageCerberus is to add a specific set of low-level image features to standard image features, able to detect the presence of obfuscated text embedded into an image, which is an indication of possible e-mail "spamminess".

In order to better evaluate the effectiveness of image spam filters and above all the robustness of OCR-based approaches against obfuscated text, in collaboration with the P.R.A.Group Of the University of Cagliari, we developed an artificial image spam generator, called

Image Spam Lab which is capable to implement several obfuscation techniques used in real spam images, with a desired degree of obfuscation.

The capabilities of our plugins can be understood from the video downloadable form this page. In particular this video show the response against clean and obfuscated ImageSpam (generated by Image Spam Lab) of three spam accounts equipped with a standard version of SpamAssassin:

• account without any image analysis

• account with BayesOCR plugin

• account with ImageCerberus plugin

These results point out that:

• standard filters are not able to detect image spam

• BayesOCR plugin allows  to detect only clean image spam

• ImageCerberus plugin allows  to detect clean and obfuscated image spam