1. 1 DSpin: Detecting Automatically Spun Content on the Web Speaker : Ting Luo 2014/05/26 Qing Zhang, David Y. Wang, Geoffrey M. Voelker University of California, San Diego Network and Distributed System Security Symposium(NDSS 2014)

2. 2 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

3. 3 Introduction Search Engine Optimization (SEO) Black Hat SEO techniques that are used to get higher search rankings in an unethical manner Spinning To generating and posting Web spam What is Spinning ? replaces words restructures original content to create new versions with similar meaning but different appearance

4. 4 Introduction Using Spinning in SEO to increase page ranks 1.create many different versions of a single seed article 2.post those versions on multiple Web sites with links pointing to a site being promoted Target Site A B C D Original

5. 5 Introduction Goal detect automatically spun content on the Web Input a set of article pages crawled from various Web sites output a set of pages flagged as automatically spun content

6. 6 Introduction Contributions 1.Spinning characterization The Best Spinner 2. Spun content detection detecting automatically spun content based upon immutables 3. Behavior of article spammers

7. 7 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

8. 8 Background And Previous Work A. Spinning Overview

9. 9 Example Both links to adult webcam sites The spun content is in English, but has been posted to German and Japanese wikis You have actually seen the feared demon-eye impact that occurs when the camera flash bounces off the eye of a person or animal You’ve seen the dreaded demon-eye impact that happens when the camera flash bounces off the eye of an individual or animal Background And Previous Work A. Spinning Overview

10. 10 (6) SPAM Content Background And Previous Work A. Spinning Overview

11. 11 Background And Previous Work B. Article Spam Detection Web spam taxonomies –content spam Quilted pages Keyword stuffing –link spam Page hijacking Link farms

12. 12 Background And Previous Work C. Near-duplicate Document Detection Near-duplicate Document –Two such documents differ from each other in a very small portion that displays advertisements Fingerprinting Algorithm –A procedure that maps an arbitrarily large data item (such as a computer file) to a much shorter bit string –reduce storage and computation costs

13. 13 Background And Previous Work C. Near-duplicate Document Detection From : http://en.wikipedia.org/wiki/Fingerprint_(computing)

14. 14 Background And Previous Work C. Near-duplicate Document Detection The classic approach - Shingles [1] –The hash value of a k-gram which is a sub-sequence of k successive words –The sets of shingles constitutes the set of features of a document Enables a graph representation for similarity among pages pages as nodes edges between two pages that share shingles above a threshold [1] Gurmeet Singh Manku, Arvind Jain, Anish Das Sarma, ‘Detecting Near-Duplicate for Web Crawling,’ 2007

15. 15 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

16. 16 The Best Spinner(TBS) A. TBS

17. 17 The Best Spinner(TBS) A. TBS A popular spinning tool –$77 per year –requires registration with a username and password synonym dictionary –requires credentials at runtime to allow the tool to download an updated version Spintax –{Home|House|Residence|Household}

18. 18 The Best Spinner(TBS) A. TBS Parameters –Frequency every word, or one in every second, third, or fourth word –Remove original removes the original word from the spintax alternatives {Home|House|Residence|Household}  {House|Residence|Household} –Auto-select inside spun text when selected, spins already spun text

19. 19 The Best Spinner(TBS) A. TBS {You can| You are able to | It is possible to | You’ll be able to | You possibly can}

20. 20 The Best Spinner(TBS) B. Reverse Engineering TBS During every startup –downloads the latest version of the synonym dictionary –Save as the file tbssf.dat in an encrypted format (base64 encoding) After Reversing Engineering TBS –use an authentication key to download the synonym dictionary Synonym dictionary –8.4 MB in size –has a total of 750,114 synonyms grouped into 92,386 lines

21. 21 The Best Spinner(TBS) B. Reverse Engineering TBS Authentication key

22. 22 The Best Spinner(TBS) C. Controlled Experiments 5-12% 6-14%

23. 23 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

24. 24 Similarity Similarity score classic Jaccard Coefficient –take all the words from the two documents, A and B –compute the set intersection over the set union across all the words

25. 25 Similarity How to compute the intersection and size of two documents? Extention A. Methods Explored B. The Immutable Method C. Verification Process

26. 26 Similarity A. Methods Explored (1)Shingling Computing shingles, or n-grams, over the entire text with a shingle size of four –a sentence “a b c d e f” is the set of three elements “a b c d”, “b c d e”, and “c d e f”. the intersection is the overlap of shingles between two documents

27. 27 Similarity A. Methods Explored low similarity between 21.1–60.7% Although useful for document similarity, it is not useful for identifying spun content given the low similarity scores

28. 28 Similarity A. Methods Explored (2) Parts-of-speech Standford NLP package –For each sentence, the NLP parser returns the original sentence with parts-of-speech tags for every word –use the parts-of-speech lists as the comparison unit

29. 29 Similarity A. Methods Explored TBS can replace single words with phrases, and phrases comprised of multiple words can be spun into a single word

30. 30 Similarity B. The Immutable Method Separate each article’s words into –mutables –Immutables focus entirely on the list of immutable words from two articles to determine if they are similar

31. 31 Similarity A. Methods Explored Ratios are above 90% for most spun content provides a clear separation between spun and non-spun content

32. 32 Similarity B. The Immutable Method Benefit –it also greatly decreases the number of bytes needed for comparison by reducing the representation of each article by an order of magnitude.

33. 33 Similarity C. Verification Process mutable verifier Steps –it sums all the words that are common between the two pages, and adds it to the total overlap count pages –It computes the synonyms of the remaining words from one page and determines if they match the words of the other page –taking the synonyms of the synonyms of the remaining words and comparing them in a similar fashion to step two

34. 34 Similarity A. Methods Explored Has a much higher overhead

35. 35 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

36. 36 Methodology A. Data Sets Wikis –purchase a Fiverr job offering to create 15,000 legitimate backlinks –Crawled the recent posts on each of the wikis 37M pages for December 2012 GoArticles –Allows users to build backlinks as “dofollow” that can affect search engine page rankings. –crawl over 1M articles posted between January 2012 to May 2013

37. 37 Methodology B. Filters Visible text –remove all pages that do not contain any visible text on the page Content tag –Wiki : div labeled “bodyContent” –GoArticles : div with “class=article” –If it lacks of this tag, then remove it

38. 38 Methodology B. Filters Word count –Discard small pages –Threshold of 50 words Link density –Discard pages with an unusually high link density Foreign text –Only evaluate the immutable method on pages with mostly English text

39. 39 Methodology C. Inverted Indexing Definition – id : a unique index corresponding to an article immu is an immutable that occurs in id. – > Each group represents all document ids that contain the immutable – the total number of immutables that overlap between id i and id j

40. 40 Methodology C. Inverted Indexing Calculate the similarity score between each two pages Set the threshold to be 75% > 2 articles

41. 41 Methodology D. Clustering graph representation –each page(ids) is a node –each pair has an edge Each connected subgraph represents a cluster

42. 42 Methodology E. Exact Duplicates and Near Duplicates Exact duplicates –Use a hash over each page (MD5 sum) –two articles are identical if their MD5 sums match Near Duplicates –Using mutable verifier –100% mutable match, but with mismatching MD5 sums

43. 43 Methodology E. Exact Duplicates and Near Duplicates For example –The English professor Synonym dictionary –{The|…} {The English professor|…} Ideal – is a mutable phrase In fact – will be marked as mutable – will be marked as immutable

44. 44 Methodology F. Hardware 24 physical nodes running Fedora Core 14 Each node has –a single Xeon X3470 Quad-Core 2.93GHz CPU and 24 GB of memory Runs on –Hadoop 1.1.2 and Pig 0.11.1 jobs

45. 45 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

46. 46 Spinner In The Wild A. Volume Wiki –68.0% as SEO spam –35.6% are spun content GoArticles has drastically less spun content (7.0%) than the wiki data set

47. 47 Spinner In The Wild B. False Positives False positives –two articles that appear in the same cluster but are unrelated Randomly sampled 99 clusters, for each one chose 2 pages. –found no evidence of false positives

48. 48 Spinner In The Wild C. Cluster Sizes Wiki data setGoArticles data set

49. 49 Spinner In The Wild D. Content most of the popular words appear to relate to sales and services

50. 50 Spinner In The Wild E. Domains 1. Spun content across domains –the average cluster spans across 12 ± 27 domains –spammers target multiple domains when posting spun content, instead of a single site

51. 51 Spinner In The Wild E. Domains It indicates a strong, positive correlation between larger scale spinning campaigns and a larger number of targeted domains

52. 52 Spinner In The Wild E. Domains 2. Spun content per domain –The bulk of the distribution are when domains have 15%–65% spun content

53. 53 Spinner In The Wild F. Timing Wiki –75% of duration <=1 day –50% of duration <=3 days

54. 54 Spinner In The Wild G. Backlinks Wiki –Links occur on 99.97%±1.41% of pages per cluster on average

55. 55 Spinner In The Wild G. Backlinks GoArticles –larger spinning campaigns generally targeting a smaller set of unique backlinks and domains than the number of pages

56. 56 Spinner In The Wild H. GoArticles as Seed Pages the majority of cross domain clusters contain many wiki pages (31.6 on average), compared with just 1.2 on average for GoArticles

57. 57 Spinner In The Wild H. GoArticles as Seed Pages

58. 58 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

59. 59 Disucssion Response of spammers –Change the dictionary frequently –tools could compute spun content remotely Future work –Other spinning tools or human-generated spun content

60. 60 Outline 1. Introduction 2. Background And Previous Work 3. The Best Spinner 4. Similarity 5. Methodology 6. Spinning In The Wild 7. Disussion 8. Conclusion

61. 61 Conclusion Proposed a method for detecting automatically spun content on the Web Implement a tool – Dspin –operates on sets of crawled Web pages to identify spun content

62. 62 Q & A