Information Revelation and Privacy in Online Social Networks

1. Information Revelation and Privacy in Online Social Networks Ralph Gross and Alessandro Acquisti rgross@cs.cmu.eduacquisti@andrew.cmu.edu Heinz Seminars, October 3rd, 2005

2. Information revelation and privacyin online social networks • Online social networks (OSN): sites that facilitate interaction between members through their self-published personal profiles • How much do users of OSN reveal about themselves online? • A lot • To whom? • Friends and strangers • Why?

3. Why? • Rationality hypothesis: signaling • Low privacy sensitivity • Herding behavior • Peer pressure • Myopic discounting • Incomplete information • …

4. Privacy, economics, and rationality • Incomplete information • Bounded rationality • Affective processes, psychological/behavioral deviations from pure rationality model

5. Our study • Starts research on privacy implications of OSN • Provides first quantification of observed behavior • Studies actual usage data • Discusses trade-offs and incentives and advances behavioral hypotheses • Yet, still preliminary • Implications extend beyond OSN domain

6. Agenda • Online social networks • The Facebook • CMU students and the Facebook • Usage data • Patterns of information revelation • Inferred privacy preferences • Risks and trade-offs • User survey (pilot) • Users’ knowledge and expectations • Drivers and incentives • Next step • Experiments

7. Online Social Networks

8. What are online social networks? • Sites that facilitate interaction between members through their self-published personal profiles • Common core: • Through the site, individuals offer representations of their sel[ves] to others to peruse, with the intention of contacting or being contacted by others, to meet new friends or dates, find new jobs, receive or provide recommendations, … • Progressive diversification and sophistication of purposes and usage patterns • Social Software Weblog groups hundreds of social networking sites in nine categories (business, common interests, dating, facetoface facilitation, friends, pets, photos, …) • Classifieds <> OSN <> blogs

9. A history of online social networks • 1960s: Plato (University of Illinois) • 1997: SixDegrees.com • After 2002: commercial explosion • Friendster, Orkut, LinkedIn, …, • Viral growth with participation expanding at rates topping 20% a month • 7 million Friendster users; 2 millions MySpace users; 16 million registered on Tickle to take personality test (Leonard 2004) • Revenues: advertising, data trading, subscriptions • Media attention: Salon, NYT, Wired, …

10. Research on online social networks • boyd (2003): trust and intimacy on OSN • Donath and boyd (2004): representation of self on OSN • Liu and Maes (2005): harvesting OSN for recommender systems • (some additional research uses OSN data for other purposes)

11. From (social) network theoryto online networks • Milgram (1967): the small world problem • Watts (2003): six degrees • Granovetter (1973, 1983): weak and strong ties • Milgram (1977): the familiar stranger • What about the “unknown buddy”?

12. Social network theory and privacy • Strahilevitz (2005): Discourse about privacy should be based “on what the parties should have expected to follow the initial disclosure of information by someone other than the defendant” • Consideration of expected information flows within/outside somebody’s social network should inform that person’s expectations for privacy • However, application to online social network reveals challenges

13. Online vs offline social networks • Offline: extremely diverse ties. Online: simplistic binary relations (boyd 2004) • Number of strong ties not significantly increased, but number of weak ties can increase substantially (Donath and boyd 2004) • From a dozen of intimate ties plus 1000 to 1700 “acquaintances,” to hundreds of direct “friends” and hundreds of thousands of relations

14. Hence: • Online social networks are vaster and have more weaker ties than offline social networks • An imagined community? • Anderson (1991) • Intimacy and trust • Sharing same personal information with a large and potential unknown number of friends and strangers • Intimate with everybody? (Gerstein 1984) • Ability to meaningfully interact with others is mildly augmented, while ability of others to access the person is significantly enlarged

15. Online social networks and personal information • Pretense of identifiability changes across different types of sites Anonymous <> Pseudonymous <> Fully identified • Type of information revealed or elicited often orbits around hobbies and interests, but can stride from there in different directions • From classified to journals • Visibility of information is highly variable • Members only • Everybody

16. Online social networks and privacy • Privacy implications of OSN depend on the level of identifiability of the information provided, its possible recipients, and its possible uses • Re-identification • Two directions: known>additional information; unknown>known • To whom may identifiable information be made available? • Site, third-parties (hackers, government), users (little control on social network and its expansion) • Risks • From identity theft to online and physical stalking; from embarrassment and blackmailing to spam and price discrimination

17. Online social networks and privacy • And yet: • OSN can also offer tools to address online privacy problems • “Social networking has the potential to create an intelligent order in the current chaos by letting you manage how public you make yourself and why and who can contact you.” Tribe.net CEO Mark Pincus • Is that true?

18. The Facebook

19. The Facebook • www.facebook.com • Started February 2004 • Attracted Silicon Valley funding • Has spread to 2000 schools and 4.2 million users • Typically attracts 80 percent of a school’s undergraduate population • Also gets graduate students, faculty members, staff, and alumni • Now targeting high schools • Growing media attention

21. Facebook‘s privacy policy • …is lax, but straightforwardly so: “Facebook also collects information about you from other sources, such as newspapers and instant messaging services. This information is gathered regardless of your use of the Web Site.” … “We use the information about you that we have collected from other sources to supplement your profile unless you specify in your privacy settings that you do not want this to be done.” … “In connection with these offerings and business operations, our service providers may have access to your personal information for use in connection with these business activities.”

22. Facebook and unique privacy issues • Unique data • Includes home location, current location (from IP address), etc. • Uniquely identified • College email account • Contact information • Ostensibly bounded community • “Shared real space” • …or imagined community?

23. CMU students and the Facebook: usage data

24. Studies • Gross and Acquisti, Proceedings of WPES 2005 • Acquisti and Gross, Proceedings of PET 2006

25. Data gathering • In June 2005, we created Facebook profiles with different characteristics • E.g., degree of connectedness, geographical location, … • We searched for CMU Facebook members’ profiles using advanced search feature and extracted profile IDs • Downloaded profiles • Inferred additional information not immediately visible from profiles

26. Demographics

27. Demographics

28. Demographics

29. Information revelation

30. Information revelation • Male users 63% more likely to leave phone number than female users • Single male users tend to report their phone numbers in even higher frequencies

31. Data verifiability

32. Data verifiability

33. Privacy risks • Stalking • Re-identification • Digital dossier

34. Privacy risks: Stalking • Real-World Stalking • College life centers around class attendance • Facebook users put home address and class list on their profiles; whereabouts are known for large portions of the day • Online stalking • Facebook profiles list AIM screennames • AIM lets users add “buddies” without notification • Unless AIM privacy settings have been changed, adversary can track when user is online

35. Privacy risks: Re-identification • Demographics re-identification • 87% of US population is uniquely identified by {gender, ZIP, date of birth} (Sweeney, 2001) • Facebook users that put this information up on their profile could link them up to outside, de-identified data sources • Face re-identification • Facebook profiles often show high quality facial images • Images can be linked to de-identified profiles on e.g. Match.com or Friendster.com using face recognition • Social Security Number re-identification • Anatomy of a social security number: xxx yy zzzz • Based on hometown and date of birth xxx and yy can be narrowed down substantially

36. Privacy risks: Digital Dossier • Users reveal sensitive information (e.g. current partners, political views) in profiles • Simple script programs allow adversaries to continuously retrieve and save all profile information • Cheap hard drives enable essentially indefinite storage

37. Privacy risks

38. Data accessibility

39. Data accessibility

40. Data accessibility • Profile Searchability • We measured the percentage of users that changed search default setting away from being searchable to everyone on the Facebook to only being searchable to CMU users • 1.2% of users (18 female, 45 male) made use of this privacy setting • Profile Visibility • We evaluated the number of CMU users that changed profile visibility by restricting access from unconnected users • Only 3 profiles (0.06%) in total fall into this category • Caveat: We would not detect users who had made themselves both unsearchable and invisible within CMU network (safe to assume their number is very low)

41. Data accessibility

42. Actual data accessibility:An imagined community? • Extensive, uncontrolled social networks • Fragile protection: • Fake email addresses • Manipulating users • Geographical location • Advanced search features • Using advanced search features various profile information can be searched for, e.g. relationship status, phone number, sexual preferences, political views and (college) residence • By keeping track of the profile IDs returned in the different searches a significant portion of the previously inaccessible information can be reconstructed • AIM • Facebook profiles are, effectively, public data

43. Actual data accessibility:An imagined community • “What a great illustration of how things you might not mind being public in one context can cause all sorts of problems when they wind up globally public.” • CMU student

44. Initial hypotheses • Default settings (Mackay 1991)/ Myopic discounting? • Less than 2% make their profiles less searchable • Less than 1% make their profiles less visible • Peer pressure • Incomplete information and biased perspectives • An imagined community • Or simply: • Low privacy concerns • Signaling • Single males list phone number with highly significant more frequency than females

45. User survey (pilot)

46. (Pilot) Survey • Goals • Understand CMU Facebook’s users degree of awareness about the site and its information revelation patterns; understand their privacy attitudes and expectations • Thirty-six online questions • Anonymous, paid • Pilot • 50 subjects • Focused on Facebook users • Survey link

47. CAVEAT: The following results are based on our pilot test (50 subjects). Hence they must only be considered suggestive trends rather than robust evidence. We are now exploring the same questions in the full survey – please contact us for the most recent results: acquisti@andrew.cmu.edu.

48. Generic concerns (7-point Likert scale)

49. Specific concerns (7-point Likert scale)

50. Attitudes vs. behavior • Share of users with high sensitivity (Likert >5) to partner/sexual orientation information who provide it on Facebook: ~70% • Share of users with high sensitivity (Likert >5) to home location and class schedule information who provide it on Facebook: ~32% • Share of users with high sensitivity (Likert >5) to contact information who provide it on Facebook: ~42%