1. Reporter: Li, Fong Ruei National Taiwan University of Science and Technology 9/19/2015Slide 1 (of 32)

2.  Large-Scale Automatic Classification of Phishing Pages, Colin Whittaker, Brian Ryner, Marria Nazif, NDSS '10, 2010 9/19/2015Slide 2 (of 32)

3.  Introduction  Phishing Classifier Infrastructure  Evaluation  Conclusion 9/19/2015Slide 3 (of 32)

4.  Phishing is form of identity theft  social engineering techniques  sophisticated attack vectors  To harvest financial information from unsuspecting consumers.  Often a phisher tries to lure her victim into clicking a URL pointing to a rogue page. 9/19/2015Slide 4 (of 32)

5.  Overall System Design  Our system classifies web pages submitted by end users and collected from Gmail’s spam filters.  These features describe the composition ▪ the web page’s URL ▪ the hosting of the page ▪ the page’s HTML content as collected by a crawler 9/19/2015Slide 5 (of 32)

6.  Classification Workflow  The first process extracts features about the URL of the page.  The second process obtains domain information about the page and crawls it  The final process assigns the page a score based on the collected features representing the probability that the page is phishing 9/19/2015Slide 6 (of 32)

7.  Candidate URL Collection  We receive new potential phishing URLs in reports ▪ from users of our blacklist ▪ from spam messages collected by Gmail 9/19/2015Slide 7 (of 32)

8.  URL Feature Extraction  The first process in the workflow, the URL Feature Extractor, looks only at the URL of the page to determine features.  If it matches a whitelist of high profile, safe sites, then the URL Feature Extractor drops the URL from the workflow entirely.  We manually compile this whitelist of 2778 sites 9/19/2015Slide 8 (of 32)

9.  URL Feature Extraction  One feature this process extracts is whether the URL contains an IP address for its hostname. 9/19/2015Slide 9 (of 32)

10.  URL Feature Extraction ▪ Another feature this process extracts is whether the page has many host components ▪ Phishers commonly use a long hostname, prepending an authentic-sounding host to their fixed domain name, to confuse viewers into believing that the page is legitimate. 9/19/2015Slide 10 (of 32)

11.  URL Feature Extraction  Phishers often include characteristic strings in their URLs to mislead viewers.  These can include the trademarks of the phishing target, like “abbeynational” in the example above, or more general phrases associated with phishing targets, like “login”.  The feature extractor transforms each of these tokens into a boolean feature, such as “The path contains the token ‘login.’” 9/19/2015Slide 11 (of 32)

12.  Fetching Page Content  The URL Feature Extractor also collects URL metadata, including PageRank, from Google proprietary infrastructure  We also use a domain reputation score computed by the Gmail anti spam system as a feature. ▪ This score is roughly the percentage of emails from a domain which are not spam 9/19/2015Slide 12 (of 32)

13.  Hosting and Page Feature Extraction  The Content Fetcher process crawls the page and gathers its hosting information. ▪ It records the returned IPs, name servers, and name server IPs. ▪ It also geo locates these IPs, recording the city, region, and country 9/19/2015Slide 13 (of 32)

14. Machine Learning and Bioinformatics Laboratory  Hosting and Page Feature Extraction  The Content Fetcher sends the URL to a pool of headless web browsers to render the page content.  After the browser renders the page, the Content Fetcher receives and records the page HTML, as well as all iframe, image, and javascript content embedded in the page 9/19/2015Slide 14 (of 32)

15. Machine Learning and Bioinformatics Laboratory  Page Classification  To compute the score for the page in log odds, the classifier combines these values using a logistic regression  The score translates to the computed probability that the page is phishing 9/19/2015Slide 15 (of 32)

16.  Page Classification  Before the classifier automatically blacklists the page, it checks to make sure that the page does not have a high PageRank 9/19/2015Slide 16 (of 32)

17.  Evaluation Dataset  First ▪ contains data collected between April 16, 2009 and July 14, 2009 with labes from July 15, 2009. ▪ examine our selected features and train our evaluation models  Second ▪ collected during the first two weeks of August, 2009, as a validation dataset. 9/19/2015Slide 17 (of 32)

18. 9/19/2015Slide 18 (of 32)

19. 9/19/2015Slide 19 (of 32)

20. 9/19/2015Slide 20 (of 32)

21. 9/19/2015Slide 21 (of 32)

22. Machine Learning and Bioinformatics Laboratory 9/19/2015Slide 22 (of 32)

23. Machine Learning and Bioinformatics Laboratory 9/19/2015Slide 23 (of 32)

24. 9/19/2015Slide 24 (of 32)

25. 9/19/2015Slide 25 (of 32)

26. 9/19/2015Slide 26 (of 32)

27. 9/19/2015Slide 27 (of 32)

28. 9/19/2015Slide 28 (of 32)

29. 9/19/2015Slide 29 (of 32)

30. 9/19/2015Slide 30 (of 32)

31.  we describe our large-scale system for automatically classifying phishing pages which maintains a false positive rate below 0.1%.  Our classification system examines millions of potential phishing pages daily in a fraction of the time of a manual review process 9/19/2015Slide 31 (of 32)

32. 9/19/2015Slide 32 (of 32)