User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability

1. User Identities across Social Networks: Quantifying Linkability and Nudging Users to control Linkability Dr. Ponnurangam Kumaraguru (Advisor)Srishti Chandok

2. Thesis Committee Dr. Arun Balaji Buduru, IIIT-‐Delhi Dr. Anuja Arora, JIIT-‐Noida Dr. Ponnurangam Kumaraguru (Chair), IIIT-‐Delhi 2

3. Publication Our work is accepted as a short paper / 8 pages at Social Informatics 2017. 3

4. Why people join multiple OSNs??? Type of content being shared: ✦ Images ✦ Videos ✦ Short messages ✦ Combination of messages, video and images ✦ Online reviews ✦ Discussion forums Type of network being offered: ✦ Professional network ✦ Personal network 4

5. Notion of Linkability Linkability is a metric which quantifies closeness between two identities belonging to the same user on different social networks 5 Username: RaineRamirez1 Name: Chris Raine Ramirez Location: Caloocan City Website: NULL Username: Rainevouz Name: Christopher Delgar Ramirez Bakunawa Location: San Jose del Monte, Bulacan Website: NULL 0.31Linkability Score = There is a 31% chance that Rainevouz & RaineRamirez1 is the same person

6. Motivation 6 ✦ Social audience = 437,632 + 153,000 + 805,097 or less?? ✦ Targeted Marketing using aggregated data De-duplicating audience - finding linkability across OSNs

7. Motivation 6 Social Engineering - Aggregation of information makes attacking easy

9. 9 1. Hacker studies Karla who is active on 2. Information aggregation: • Likes playing guitar and softball • Works on making reports at the office and recently received an award • Likes wine • Nick is her boss and Curt is her colleague 3. Hacker crafts an email 4. Downloaded on Karla’s machine 5. Hacker installs a remote access tool on the machine

10. Primary School name on 10 Motivation Streets close to ‘xyz’ school ‘xyz’ school Security Question: Street where you grew up? Cracking passwords - Personal data across multiple social networks to gather answers for password recovery questions

11. Research Aim “ Develop a real-‐time system which can help users to maintain their linkability across social networks. ” AIM Linkability Score Computation Linkability Nudge Design 11

12. Current State of Art Identity Resolution 12 Privacy Nudges Linkability Score Linkability Nudge

13. Current State of Art Identity Resolution 13 Privacy Nudges

14. Identity Resolution(Contd..) Identity Resolution: ✦ MOBIUS - ACM, SIGKDD (2013), 91% accuracy ✦ NEMO - ACM, HyperText (2015), 41% accuracy ✦ HYDRA - ACM, SIGMOD (2014) 14

15. Current State of Art Identity Resolution 15 Privacy Nudges

16. Privacy Nudges (Contd..) Privacy Nudges: ✦Wang, Yang, et al. Privacy nudges for social media: an exploratory Facebook study - Profile Picture nudge, timer nudge and sentiment nudge 16

17. 17 ✦Ronald. Context is everything sociality and privacy in online social network sites - Segregation of audience for profile attributes of user on OSNs so that its visibility is controllable. Privacy Nudges (Contd..)

18. Novelty 18 ‹#› Nudging users to prevent disclosures owing to the resolution of their multiple identities Identity Resolution Privacy Nudges

19. Architecture Diagram Linkability Score Linkability score exceeds range? Linkability Nudge NO YES 19 Activities performed on OSNs Recomputation

20. Linkability Score Linkability Score Linkability score exceeds range? Linkability Nudge NO YES 20 Activities performed on OSNs Recomputation 1. Baseline Methods Weighted Sum Probabilistic 2. Reformed Linkability Score

22. Baseline Methods 1. Data Collection 2. Methods 22 Weighted Sum Method Probabilistic Method

23. Data Collection 23

24. Data Collection (Contd..) 24 Streaming API Tweets Fb.me links Link Expander Database http://fb.me/8dR49RHpQ https://www.facebook.com/ christie.andresen/posts/ 10210420711856356

25. Description of Data Count Positive Data 23,985 Negative Data (Type I) 96,130 Negative Data (Type II) 24,560 Positive Data: <IFb> = <ITw>, identities are same [bob12, bob12] Negative Data (Type I): <IFb> ≠ <ITw> but the identities appear to be similar [bob_c, bob_d] Negative Data (Type II):> <IFb> ≠ <ITw> and the identities appear to be dissimilar [bob, alice] Data Collection (Contd..) 25

26. Weighted Sum Method Weighted Sum Calculator 26 Feature Extractor Metric Calculator Linkability Score Linkability = ∑wi * fi Score ∑wi

27. Weighted Sum Method Username Name Location Website Longest Common Subsequence, Edit Distance, etc Weighted Sum Calculator 0.31 27 Username = 0.39 Name = 0.28 Location = 0.6 Website = 0 Username: RaineRamirez1 Name: Chris Raine Ramirez Location: Caloocan City Website: NULL Username: Rainevouz Name: Christopher Delgar Ramirez Bakunawa Location: San Jose del Monte, Bulacan Website: NULL Feature Extractor Metric Calculator Linkability Score Linkability = ∑wi * fi Score ∑wi

28. Need for different Weights Feature weights = 1,1,1,1 for username, name, geo-location and website/url, respectively. Feature weights = 2,3,4,1 for username, name, geo-location and website/url, respectively. To increase the difference between positive data and negative data thereby, ensuring that negative data (Type I) identity would not be mistaken to be as positive identity. 28

29. 29 Feature Extractor Metric Calculator Linkability Score Probability Finder Linkability = Pr(pd) Score [Pr(pd) + Pr(nd1) + Pr(nd2) ] Probabilistic Method

30. Probabilistic Method 30 Username Name Location Website Longest Common Subsequence, Edit Distance, etc Probability Finder 0.12 Username = 0.39 Name = 0.28 Location = 0.6 Website = 0 Username: RaineRamirez1 Name: Chris Raine Ramirez Location: Caloocan City Website: NULL Username: Rainevouz Name: Christopher Delgar Ramirez Bakunawa Location: San Jose del Monte, Bulacan Website: NULL Feature Extractor Metric Calculator Linkability Score Linkability = Pr(pd) Score [Pr(pd) + Pr(nd1) + Pr(nd2) ]

31. Comparison of Baseline methods Accuracy is 87% when threshold value = 0.39 [feature weights used are 2, 3, 4 and 1 for username, name of user, location and website features, respectively] Accuracy is 32% when threshold value = 0.71 31

32. Limitations of Baseline Methods Probabilistic method did not produce anticipated results with quiet low accuracy. Both the methods employ a small set of features namely name, username, geo-‐location and website. They fail to capture user’s content sharing behavior. 32

33. Take-aways from Baseline Methods Weighted Sum method performed better than Probabilistic method. We will enhance the feature set using well known identity resolution techniques + our proposed Weighted Sum method to compute linkability scores 33

35. Reformed Linkability Score Leverages features from three state-of-the-art identity resolution techniques and uses Weighted Sum method to calculate linkability scores: MOBIUS NEMO HYDRA MOBIUS: http://www.public.asu.edu/~huanliu/papers/kdd2013.pdf NEMO: http://precog.iiitd.edu.in/Publications_files/19wole01-jain.pdf HYDRA: http://ink.library.smu.edu.sg/cgi/viewcontent.cgi?article=3650&context=sis_research 35

36. Reformed Linkability Score 36 Authorize And Fetch Data Feature Extraction NEMO Linkability Score MOBIUS HYDRA

37. MOBIUS 1 = srishti.chandok srishti.chandok SrishtiChandok 37

38. NEMO Srishti.chandok & SrishtiChandok Srishti Chandok & Srishti Chandok New Delhi, India & Delhi, India & & 1 0.75, 0 1 0.98 1 0.79 38 Username Similarity Score Name Similarity Score Location Similarity Score Profile Image Similarity Score Content Similarity Score & Weighted Sum

39. HYDRA Srishti Chandok & Srishti Chandok 'Salwan Public School, Rajinder Nagar, New Delhi 110060', 'I P University', 'maharaja agarsen institute of technology', 'IIIT Delhi’ & MTech from IIIT Delhi, BTech from Maharaja Agrasen Institute of Technology, IPU, SPS IIIT Delhi Teaching Assistant & MTech from IIIT Delhi, BTech from Maharaja Agrasen Institute of Technology, IPU, SPS http://www.twitter.com/SrishtiChandok & twitter.com/srishtichandok & 39 0.32 0.24 1 1 1 0.5 Name Similarity Score Education Similarity Score Profession Similarity Score Website Similarity Score Content Similarity Score Weighted Sum

40. Linkability Nudge 40 Linkability Score Linkability score exceeds range? Linkability Nudge NO YES Activities performed on OSNs Recomputation 1. Baseline Methods Weighted Sum Probabilistic 2. Reformed Linkability Score

41. Linkability Nudge 41 Soft paternalistic intervention Alerts users whenever user behavior leads to change in linkability score beyond pre-configured range

42. Components of Linkability Nudge Browser Extension Nudge Server Linkability Compute Server 42

43. 1. Browser Extension Maintains user's identity across the entire user session. Captures user's posting activity and changes in profile attributes on all configured OSM platforms. Displays linkability nudge in various forms (notifications and color). 43 Downloads the Chrome browser extension Nudge Server Send User’s Activity Information Linkability Score and Piecharts Browser Extension

44. 2. Nudge Server Intermediary between the browser extension and linkability compute server. Receives user's access token from browser extension and sends them to linkability compute server to obtain user's data. Passes the information pertaining to user's activities like making a post or changing profile attribute to the linkability compute server. Sends across the newly computed linkability scores to the browser extension from time to time based upon user's activities. 44 Access Token for various OSNs Forward Access Token for various OSNs Linkability Score Forward User’s Activity Information Nudge Server Linkability Compute Server Browser Extension

45. 3. Linkability Compute Server Fetches user's data from the API endpoints. Implements the identity resolution methods to compute linkability scores. Receives every user’s activity information (post or profile attribute), recomputes linkability scores and sends them back to nudge server. 45 Nudge Server Linkability Compute Server Identity Resolution Algorithms NEMO | HYDRA | MOBIUS Fetch user’s data Linkability Score Linkability Score

46. Nudge Design Content-driven Color Nudge - ✦ Similar Post --> Red color nudge ✦ Dissimilar Post --> Green color nudge Attribute-‐driven Notification Nudge -‐ ✦ Profile attribute update -> Linkability Score crosses range -> Notification Popup nudge 46

47. Content-driven Color Nudge 47

48. Attribute-driven Notification Nudge 48

49. Demo 49

50. User Evaluation of the Nudge 50 Control Period Treatment Period No exposure to linkability nudge Tasks (Post and profile updates) Exposure to linkability nudge Tasks (Post and profile updates)

51. Analysis of User Evaluation 58% of the participants: Understood the broad concept of linkability score 42% of participants: More aware about the linkability of their multiple identities across OSNs 84% of the participants: Noticed the factors contributing to their linkability scores 83% of the participants: Liked Color nudge and pie-‐charts more Activities performed by one of the participants 51

52. Limitations Time delay (2-‐5 seconds) while making post during treatment period. Used uniform weights for computing linkability scores. System works for three social networks. Evaluated the nudge with a small number of participants. 52

53. Conclusions Leverage features from well known methods for identity resolution (NEM, HYDRA and MOBIUS) and use the proposed baseline method, Weighted Sum, to compute the linkability scores Identify the factors (profile attributes and content) that have contributed to the computed linkability score Design and develop linkability nudge, a soft intervention which alerts users whenever user behavior leads to change in linkability score beyond preconfigured range Perform a detailed user study in a controlled lab experiment setting to assess effectiveness and utility of proposed linkability nudge 53

54. Acknowledgements Rishabh Kaushal, PhD, IIIT-‐Delhi Committee Members Sonal and Sonu, Precogers Precog members, family and friends 54

55. Thank You! 55

56. References https://www.washingtonpost.com/investigations/social-‐engineering-‐using-‐ social-‐media-‐to-‐launch-‐a-‐cyberattack/2012/09/26/a282c6be-‐0837-‐11e2-‐ a10c-‐fa5a255a9258_graphic.html?utm_term=.148a1a0244af https://www.infosecurity-‐magazine.com/news/phishing-‐and-‐social-‐ engineering/ https://www.nytimes.com/2014/11/11/world/europe/for-‐guccifer-‐hacking-‐ was-‐easy-‐prison-‐is-‐hard-‐.html http://www.securityweek.com/social-‐media-‐makes-‐way-‐social-‐engineering http://www.propertycasualty360.com/2017/07/04/how-‐social-‐engineering-‐ fueled-‐the-‐cyber-‐attack-‐bus 56

57. Appendix 57

58. Feature Name Metrics username Hamming Distance, Longest Common Subsequence, Edit Distance, Cosine Distance, Jaccard Distance, Jaro Winkler Distance name Length of Common Substring, Length of Common Prefix & Common Suffix location Length of Common Substring, Geo-location (LAtitude & Longitude) website Canonical URL matching Features and Metrics for Baseline Methods 58

59. NEMO 59

60. HYDRA 60

User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability

User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability

Recommended

More Related Content

Similar to User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability

Similar to User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability(20)

More from IIIT Hyderabad

More from IIIT Hyderabad (20)

Recently uploaded

Recently uploaded(20)

User Identities Across Social Networks: Quantifying Linkability and Nudging Users to control Linkability