Empirical Models of Privacy in Location Sharing Eran Toch, Justin Cranshaw, Paul Hankes-Drielsma, Janice Y. Tsai, Patrick Gage Kelley, James Springfield, Lorrie Cranor, Jason Hong, Norman Sadeh Carnegie Mellon
Jan 27, 2015
Empirical Models of Privacy in Location
Sharing
Eran Toch, Justin Cranshaw, Paul Hankes-Drielsma, Janice Y. Tsai, Patrick Gage Kelley, James Springfield, Lorrie Cranor,
Jason Hong, Norman Sadeh
Carnegie Mellon
(1) Motivation
Ubicomp 2010 Carnegie Mellon
Motivation
4Ubicomp 2010 Carnegie Mellon
Privacy
‣ Location sharing applications can reveal sensitive locations (e.g., home,) the activity of the user, social encounters etc...
‣ Privacy is a major concern that may limit adoption (Tsai et al. 2009.)
by Frank Groeneveld, Barry Borsboom and Boy van Amstel.
Ubicomp 2010 Carnegie Mellon
Background
‣ Privacy
‣ Khalil and Connelly (2006)
‣ Anthony et al. (2007)
‣ Benisch et al. (2010)
Location and Mobility
‣ Eagle et al. (2006)
‣ Gonz´alez et al. (2008)
‣ Mancini et al. (2009)
‣ Cranshaw et al., 2010
Our question: What are the privacy preferences associated
with locations and mobility patterns?
7Ubicomp 2010 Carnegie Mellon
Agenda
‣ Locaccino
‣ Study
‣ Results
‣ Conclusions
(2) Locaccino
Ubicomp 2010 Carnegie Mellon
Locaccino
‣ Location sharing application
‣ Expressive privacy controls
‣ Background location tracking
‣ Research framework
10
Ubicomp 2010 Carnegie Mellon
Locators
‣ Background location reporting every 2-10 minutes, depending on movement
‣ On laptops: Location WiFi positioning by Skyhook
‣ On smartphones: WiFi positioning + GPS
For Mac and Windows
Ubicomp 2010 Carnegie Mellon
Setting Privacy Policy
Ubicomp 2010 Carnegie Mellon
Requesting Locations
(3) Study
Ubicomp 2010 Carnegie Mellon
Study‣ 28 primary participants were recruited using flyers
scattered around the Carnegie Mellon Campus and mailing list posting. They were compensated at $30 + data plan.
‣ 373 secondary participants had joined by invitation of primary participants. They were not compensated.
‣ 230 of them installed a locator, and were requested by other participants.
1. Answering Entrance Survey
3.Installing locator4.Setting up
privacy policy5.Inviting friends
3. Using Locaccin
o
4. Answering
Place Survey +
Exit Survey
2. Randomly assigned a
locator
Ubicomp 2010 Carnegie Mellon
Population and Limitation
‣ All participants are from the university community.
‣ 17 graduate students, 9 undergraduate students and 2 staff members.
‣ The study was conducted in a single city (Pittsburgh.)
‣ And in the course of a single summer month.
(4) Results
Ubicomp 2010 Carnegie Mellon
Location Entropy‣ Entropy is a measure for the
diversity of visitors to a place (Cranshaw et al., 2010)
‣ Borrowed from bio-diversity, it assigns high values to places visited by many users in equal proportions.
‣ Let p(u,l) be the observations of a user u in a location l. Entropy is defined as:
High entropy (5+)
Medium entropy (1-5)Low entropy (1)
Locations are defined based a 100m radius
Ubicomp 2010 Carnegie Mellon
Place Survey
Ubicomp 2010 Carnegie Mellon
Entropy vs. Comfort in sharing locations
Users were more comfortable sharing high entropy locations.ANOVA, friends: F=5.46 p=0.02, distant relations: F = 15.57 p=0.001
The correlation is stronger for distant social relations than with close social relations
Ubicomp 2010 Carnegie Mellon
Sharing by Place Type
Tags were grouped by a team of 3 judges to 8 categories
For distant relations
Ubicomp 2010 Carnegie Mellon
Privacy and Mobility• Visible mobility is
correlated with the number of request for the user (ANOVA: F = 14.713 p = 0.00079)
‣ High mobility users were requested twice as much as low mobility users.
‣ Number of friends and the users’ activity are non significant.
High mobilit
y users
Low mobilit
y users
Visible mobilityNumber of unique daily
locations
Median: 3.4
Ubicomp 2010 Carnegie Mellon
Requests over time
The request rate for high mobility users increased twofold over the course of the study
Ubicomp 2010 Carnegie Mellon
Privacy and Mobility
Item ANOVA F ANOVA P-value
Expressiveness (number of policy restrictions)
5.63 0.025
Number of privacy policy updates 10.75 0.0028
Correlation between visible mobility and privacy properties
High mobility users were 4 times as likely to use location restrictions and 7 times more likely to use time restrictions
24
Ubicomp 2010 Carnegie Mellon
Rule Examples
Ubicomp 2010 Carnegie Mellon
Survey Results
Item Average ANOVA F ANOVA P-value
Overall Usefulness 4.74 4.54 0.043
Friends rules usefulness 5.48 4.68 0.04
Time rules usefulness 4.74 5.14 0.03
Location rules usefulness 5.14 4.15 0.052
‣Correlation between visible mobility and survey results
7-point Likert (1 stands for not useful and 7 for very useful)
(4) Conclusions
Ubicomp 2010 Carnegie Mellon
Conclusions
‣ Some privacy preferences can be predicted by location entropy and mobility.
‣ Enhancing location sharing: by suggesting helpful defaults, checking-in in high entropy places etc.
‣ Establishing privacy sensitive location reporting for location aware systems.
‣ Other fields? Is entropy related to other phenomena? Check Session VII
‣ Lots of future work...
Thank you
More info:
http://www.cs.cmu.edu/~eran/
Carnegie Mellon
Locaccino demo - tomorrow’s
lunch
Ubicomp 2010 Carnegie Mellon
Location Privacy Preferences
‣Which measure best predicts the location privacy preferences?
ANOVA p-value
Measure friends and family
distant relations
Number of unique visitors
0.48 0.3
Number of observations 0.17 0.001
User’s visits to the location
0.98 0.22
Location entropy 0.02 0.001
30
Ubicomp 2010 Carnegie Mellon
Statistics
Item Average
Number of friends 12.86
Number of location observations 1,417,095