the promise of differentially private social network analysis · picture from andreas...
TRANSCRIPT
The Promise of Differentially PrivateSocial Network Analysis
Vishesh Karwa
Carnegie Mellon University
LARC – March 27th 2015
What is Statistical Privacy?
Overview of Privacy Research
Sharing Networks data privately
In This Talk
Why is Privacy Important?
7
Three methods for sharing data
www.sangrea.net
MechanismDatabase of
People/RelationsUsers
f(G)
queries
answers
Government,researchers,businesses
(or) Maliciousadversary
Privacy in Statistical Databases
8
Why is Privacy Hard?
Why is Network Privacy Hard?
Attacks on Past Techniques
Picture from Andreas Haeberlen’s slides
Netflix attack [Narayanan, Shmatikov 2008]
Picture Courtesy – Adam Smith and Arvind Narayanan
Lessons Learned
In This Talk
The Cryptographic Solution to Privacy
P(Z|X) P(Z|X’)
x x’
Edge Differential Privacy*
x
x’
P(.|x) P(.|x’)
Differential Privacy - Properties
I am okay with giving my data for the study, but I need to protect
my privacy.
Don’t worry, no one will learn anything more about you than what they already know.
The Differential Privacy Guarantee
How to achieve Differential Privacy?
Global Sensitivity:
Example - Laplace Mechanism
20
Laplace Mechanism:
f(G) f(G’)
In This Talk
Key issues with Differential privacy*
G
G
An ERGM framework for networks
The Beta model
1
2
5
3
4
Private estimator of beta model
Step 1 - Release degree sequence
Step 2 - Re-estimate Degree Sequence
Step 3 - Estimate parameters
Karate Data Set
Likoma n=250, m = 248 Degree sequence of people on Likoma Island
Karate n = 34, m = 78 Network of Members of Karate club
Likoma Island
Likoma n=250, m = 248 Degree sequence of people on Likoma Island
Karate n = 34, m = 78 Network of Members of Karate club
More general ERGMs…
Randomized ResponseOld Wine in new Bottle
Inference with Randomized data
Approximate Likelihood Inference
Faux Mesa High
Teenage friendship study
KL divergence
Faux Mesa High
Teenage Friendship Data
Summary
Thanks!
Questions?