| Phishing, the mass distribution of spoofed email messages with return addresses, links, and branding that appear to come from banks, insurance agencies, retailers, or credit card companies, has emerged as the latest threat for the theft of personal information over the Internet. It has the potential of inflicting serious financial losses to individuals who fall victim to phishing attacks. Beyond the direct costs of individual phishing attacks, this Internet-based fraud is eroding customer confidence in the online operations of institutions targeted by such attacks. This impediment to customer utilization of online resources undermines the ability of companies and organizations to roll out Web-based applications for customers and to utilize the highly efficient communications methods that email and the Web represent. This trend is also not limited to the direct targets of phishing. As this problem goes unsolved, all organizations utilizing on-line e-commerce tools are affected by the rising distrust of Internet-based communications and transactions. The thesis solves this problem through the use of machine learning techniques, models and algorithms. In particular, we detect phishing emails before they ever get to the mailboxes of unsuspecting users. The thesis makes the following contributions to solve the phishing problem: characterizing and formalizing phishing from spamming, providing the first publicly available and massive data set on phishing for future research on phishing, using machine learning approaches that identify authorship, embedding domain knowledge that asks and answers questions (who sends the emails, where do the links in the email direct readers to? and what do the email messages contain?), extracting significant features required to build efficient and reliable phishing filter that is computationally efficient and performs well. |
