Spam mail has long been the problem in the recent years and it affects every email users and providers. “A report in May 2009 by Symantec suggests that 90.4% of emails were spam, Google (postini) rated spam volume to be around 90-95% in all four quarters of 2009, while a report from Microsoft in September 2009 rated spam to be at 97.3% of emails.” (Feng Qian1 Abhinav Pathak2 Y. Charlie Hu2 Z. Morley Mao1 Yinglian Xie3, 2010).

There are lots of spam mail detection techniques. Currently, the statistical machine learning filter is the popular one of many commercial anti-spam filters (Seongwook Youn & Dennis McLeod, 2009). There are three types of the machine learning filter including supervised machine learning filter, semi-supervised machine learning filter and unsupervised machine learning filter. Apart from the machine learning approach, the word filters and rule-based scoring system can also detect the spam mails by simply identifying any email that contains certain key words.

There are limitations for the existing solutions. False positive is one of the examples, in the spam mail detection, means that when the mail is classified as a spam which is a positive result for the spam mail detection, but the mail is actually a legitimate message which indicates that the result is false. False positive is possible to happen for all spam filters and it is unacceptable because an important e-mail may be identified as a spam and is rejected. Besides, the rule-based methods are difficult to evaluate and need to be maintained manually. If the spammers know the keywords used for filter the spam mail, they can avoid using the keywords to keep their spam mails up to date.

The current spam mail filters need to keep up with the spam techniques. As the spam mails keep evolving, the spam mail senders frequently change the terms they used in the spam mails to bypass the keyword rule-based scoring filters. Also, they may insert some non-spam content from the book or newspaper into their mail in order to make their mails seem like legitimate mails.

We try to develop a more robust method by analyzing the spamming behaviours and mail senders' intentions that are infrequently changed. Our method includes a theorectical part and a implementation part. The theorectical part is to build a model for the spam mail senders' self-representation. The implementation part is to develop a spam mail detecting system which detects the spam mail based on the theorecticl model.

In this project, we investigate a set of spam mails to find out how the spam mail senders present themselves to deceive and persuade the mail receivers. A model containing the hyper-personal effect properties that appears in the spam mails is produced. The features in this model are the personalities and images projected by the spam mail senders when they are able to selectively present themselves over the email. These images and personalities are extracted from the set of 300 spam mails used in this project. Using a different set of spam mails may produce a different model.

Based on our model, more than one image or personality can appears in the same mail.

There is a potential for finding out some co-occurrence of the images and personalities that is significant. In this project we group the similar images and personalities into the same class as shown below: