1. Mining Privacy Settings to Find Optimal Privacy-Utility Tradeoffs for Social Network Services
Shumin Guo, Keke Chen
Data Intensive Analysis and Computing (DIAC) Lab
Kno.e.sis Center
Wright State University

2. Outline Introduction Our modeling methods The experiments Conclusions
Background
Research goals
Contributions
Our modeling methods
The IRT model
Our research hypothesis
Modeling social network privacy and utility
The weighted/personalized utility model
Trade off between privacy and utility
The experiments
Social network data from Facebook
Experimental Results
Conclusions

3. Introduction

4. Background Social network services (SNS) are popular
SNS are filled up with private info  privacy risks
Online identity theft
Insurance discrimination …
Protecting SNS privacy is complicated
Many new young users
Do not realize privacy risks
Do not know how to protect their privacy
Privacy settings
consist of tens of options
involve implicit privacy-utility tradeoff
A privacy guidance for new young users?

5. Some facts Privacy settings of Facebook 27 items
Each item is set to one of the four levels of exposure (“me only”, “friends only”, “friends of friends”, “everyone”)
By default, most items are set to the highest exposure level
the best interest to the SNS provider is to get people exposed and connected to each other

6. Research goals Understand the SNS privacy problem
The level of “privacy sensitivity” for each personal item
Quantification of privacy
The balance between privacy and SNS utility
Enhancement of SNS privacy
How to help users express their privacy concerns?
How to help users automate the privacy configuration with utility preference in mind?

7. Our contributions
Develop a privacy quantification framework that considers both privacy and utility
Understand common users’ privacy concerns
Help users achieve optimal privacy settings based on their utility preferences
We study the framework with real data obtained from Facebook

8. Modeling SNS Users’ Privacy Concerns

9. Basic idea
Use the Item Response Theory (IRT) model to understand existing SNS users’ privacy settings
Derive the quantification of privacy concern with the privacy IRT model
Map a new user’s privacy concern to the IRT model  find the best privacy setting

10. The Item Response Theory(IRT) model
A classic model used in standard test evaluation
Example, estimate the ability level of an examinee based on his/her answers to the a number of questions

11. The two-parametric model
The black curve represents Question 1 and the red Question 2.
X-axis represents the quantification of “ability” – or attitudes such as “privacy concern”
Y-axis represents the probability of giving the right answer for that question
Beta represents the difficulty level – beta2 is larger than beta1, which means Q2 is more difficult; at the same ability level, the probability of giving the right answer for Q2 is lower than Q1
Alpha represents the discrimination level of the question – whether it can clearly discriminate low-ability users from high ability ones; flat curves (alpha is small) certainly have lower discrimination power
The IRT model is learned from users’ answers to a bunch of questions (or the privacy settings for a number of items)
The two-parametric model
α  level of discrimination for a certain question
β  Level of difficulty for a certain question
θ  Level of a person’s certain trait

12. Mapping to privacy problem
Question answer  profile item setting
Ability level of privacy concern
Beta  sensitivity of profile item
Alpha  contribution to overall privacy concern

13. What we get… Probability of hiding the item relationships network
Level of privacy concern 
Probability of hiding the item
network
relationships
Current_city

14. Our Research Approach
Observation: Users disclose some profile items while hide others
If a user believes an item is not too sensitive, he/she will disclose this item
If a user perceives an item as critical to realize his/her social utility, he/she may also disclose it
Otherwise, user will hide this item
Hypothesis: Users have some implicit balance judgment behind their SNS activities
If utility gain > privacy risk  disclose
If utility gain < privacy risk  hide

15. Modeling SNS privacy Use the two-parametric IRT model
New interpretation of the IRT model
α  profile-item weight for a user’s overall privacy concern
β  Sensitivity level of the profile item
θ  Level of a user’s privacy concern

16. The complete result looks like…

17. Finding optimal settings
Theorem:
Privacy rating at i
User settings for items: 1: hidden, 0: disclosed
Probability of hiding
the item

18. Modeling SNS utility – the same method
λ  profile-item weight for a user’s SNS utility
μ  importance level of the profile item
φ  Level of a user’s utility preference
We can derive: λ = α and μ = -β
For utility model, we have:
Exposing an item  lose privacy but gain utility of that item
is the flip of sij

19. An important result
For a specific privacy setting over theta_i Privacy rating + utility rating ≈ a constant Privacy-utility are linearly related

20. The weighted/personalized utility model
Users often have clear intention for using SN but have less knowledge on privacy
Users want to put higher utility weight on (a) certain group(s) of profile items than others
Users can assign specific weights to profile items to express his/her preference
The utility IRT model can be revised with a weighted model (skip the details here)

21. Illustration of Tradeoff between privacy and utility

22. The Experiments

23. The Real Data from Facebook
Data crawled from Facebook with two accounts
Account normal: a normal Facebook account, which has a certain number of friends
Account fake: a fake account with no friends
Data crawling steps
For the friends and “friends of friends” (FOF) of account normal, crawl the profile item visibility of each user
For the same group of users, crawl the visibility of the fake account’s FoFs’ profile items again
We have the following inference rules

24. Deriving privacy settings of users
Based on the data crawled with the FoF of the two accounts, we derive the privacy setting of a user based on the following rules
E: everyone, FoF: friends of Friends,
O:the account owner, F: Friends only

25. Experimental Results
“Cleaned # of FoF”: Shared FoFs are removed to reduce the bias in modeling

26. Note:
The items that are not filled up by the user are also treated as “hidden”.
- Some people simply ignore some items or those items have no value, e.g., “graduate school”
- it is consistent with our rationale of “disclosing/hiding” items

28. Validated with 5-fold cross-validation With p-value <0.05

29. Privacy rating real setting
For each theta_i : the level of privacy concern, there is one “privacy rating” (defined in previous slide”)
The real settings may deviate from the ideal setting
Privacy rating
real setting

30. Results of learning weighted utility model
The weighting scheme is to be studied.

31. Tradeoff between privacy and utility (unweighted)
Utility rating
Privacy rating
Few people have very high
level of privacy concern
More people tend to have
lower privacy ratings, or
implicitly higher utility ratings
The relationship is about linear,
Very few have very high privacy rating.

32. Tradeoff between privacy and weighted utility

33. Conclusion
A framework to address the tradeoff between privacy and utility
Latent trait model (IRT) is used for modeling privacy and utility
We develop a personalized utility model and a tradeoff method for users to find optimal configuration based on utility preferences
The models are validated with a large dataset crawled from Facebook