Using visual and semantic features for anti-spam filters

It is well known that Unsolicited Commercial Emails (UCE), commonly known as spam, are becoming a serious problem for email accounts of single users, small companies and large institutions. The presence of spam can seriously compromise normal user activities, forcing to navigate through mailboxes to find the relatively few interesting emails, so wasting time and bandwidth, occupying their storage space. Eventually, they have often unsuitable content (as a pornographic material advertising) that could be illegal for the minors. In this realm, different countermeasures to spam have been proposed, using regulatory or technical approach. The legislative approach doesn’t obtain the desired results. A variety of technical approach are thus implemented in different anti-spam filters currently used to detect the spam content [8]. In the past, researchers have addressed this problem as a text classification or categorization problem. However, as spammers’ techniques continue to evolve and the genre of email content becomes more and more diverse, keywords-based anti-spam approaches alone are no longer sufficient. Different techniques are used to analyze the mail text, the majority are learningbased approach. Considering the spam detection as a binary classification problem, several algorithms from learning theory field can be used, such as bayesian algorithms [5] or Support Vector Machine (SVM) [6]. These systems, using the acquired knowledge, are able to discriminate the synthetic features in order to reject the mail considered as spam. Note these approaches generally don’t take into account the semantic content of e-mails. In this paper, we propose a novel anti-spam system which utilizes visual clues, in addition to semantic analysis information in the email body, to determine whether a message is spam. Figure 1: System Architecture As shown in figure 1, we propose a system that integrates image and text semantic analysis using a set of hierarchical decision systems: in particular, the different modules of the algorithm are activated only when the previous phases are not sufficient to give a positive answer. We thus apply to the text contained into e-mails some semantic analysis and Natural Language Processing (NLP) based techniques, i.e.: (i)) Latent Dirichlet Analysis (LDA)[7], in order to define a distribution of keywords for semantically categorizing e-mails contents; (ii) Latent