Privacy-Aware Data Management in Information Networks - SIGMOD 2011 Tutorial

Privacy-Aware Data Management
in Information Networks
Michael Hay, Cornell University
Kun Liu, Yahoo! Labs
Gerome Miklau, Univ. of Massachusetts Amherst
Jian Pei, Simon Fraser University
Evimaria Terzi, Boston University
SIGMOD 2011 Tutorial 1

Romantic connections in a high school
Bearman, et al.
The structure of adolescent romantic and sexual networks.
American Journal of Sociology, 2004. (Image drawn by Newman)2

A
B
G
G
G GT
GT
GT
G
G
G
G
Figure 4 (A) Graph o
component in gang ass
outbreak, Colorado Sp
1989–91 (n = 410). (B
largest component in ga
associated STD outbrea
Springs, 1989–91 (n =
Sexual network structure
on 16 August 2008sti.bmj.comDownloaded from
Potterat, et al.
Risk network structure in the early epidemic phase of hiv transmission in colorado springs.
Sexually Transmitted Infections, 2002.
Sexual and injecting drug partners
3

Figure2.NumberofnodesinthesampleNs(),obtainedbysnowballsampling,
asafunctionofextractiondistance
forseveralchoicesofthesourcenode(solid
lines)andtheiraverage(dashedline).
Figure3.Asampleofthenetwork,showingthesource(oran
centre)nodefromwhichsamplingwasstarted,thebulknode
nodes(◦).Forsurfacenodes,whichclearlyareinthemajority,
nearestneighboursarevisibleinthesample,whiletherestar
79(http://www.njp.org/)
J. Onnela et al. !
Structure and tie strengths in mobile communication networks,
Proceedings of the National Academy of Sciences, 2007
Social ties derived from a mobile phone network
4

~600 million nodes
billions of relationships
Facebook
network
data set
complex
platform for
sharing 5

Privately managing
enterprise network data
Personal Privacy in
Online Social Networks
Data: Enterprise collects
data or observes interactions
of individuals.
Control: Enterprise controls
dissemination of information.
Goal: permit analysis of
aggregate properties; protect
facts about individuals.
Challenges: privacy for
networked data, complex
utility goals.
Data: Individuals contribute
their data thru participation
in OSN.
Control: Individuals control
their connections,
interactions, visibility.
Goal: reliable and
transparent sharing of
information.
Challenges: system
complexity, leaks thru
inference, unskilled users. 6

• Privately Managing Enterprise Network Data
• Goals, Threats, and Attacks
• Releasing transformed networks (anonymity)
• Releasing network statistics (differential privacy)
• Personal Privacy in Online Social Networks
• Understanding privacy risk
• Managing privacy controls
Outline of tutorial
60
minutes
30
minutes
7

Data model
ID Age HIV
Alice 25 Pos
Bob 19 Neg
Carol 34 Pos
Dave 45 Pos
Ed 32 Neg
Fred 28 Neg
Greg 54 Pos
Harry 49 Neg
Alice Bob Carol
Dave Ed
Fred Greg Harry
Nodes
ID1 ID2
Alice Bob
Bob Carol
Bob Dave
Bob Ed
Dave Ed
Dave Fred
Dave Greg
Ed Greg
Ed Harry
Fred Greg
Greg Harry
Edges
8

Sensitive information in networks
• Disclosing attributes
• Disclosing edges
• Disclosing properties
• node degree, clustering, etc.
• properties of neighbors (e.g. mostly friends with republicans)
9

Goals in analyzing networks
• Properties of the degree
distribution
• Motif analysis
• Community structure
• Processes on networks:
routing, rumors, infection
• Resiliency / robustness
• Homophily
• Correlation / causation
Can we permit analysts to study networks without
revealing sensitive information about participants?
Example analyses
10

Naive anonymization
Original network
Naive
Anonymization
DATA OWNER ANALYST
Naive anonymization is a transformation of the network in which
identiﬁers are replaced with random numbers.
Alice Bob Carol
Dave Ed
Fred Greg Harry
4
2
5
13
6
7
8
Naively anonymized network
Good utility: output is isomorphic to the original network
Alice
Bob
Carol
Dave
Ed
Fred
Greg
Harry
6
8
5
7
2
3
4
1
11

Protection under naive anonymization
• Two primary threats:
• Node re-identification: adversary is able to deduce that node x
in the naively anonymized network corresponds to an identified
individual Alice in the hidden network.
• Edge disclosure: adversary is able to deduce that two identified
individuals Alice and Bob are connected in the hidden network.
• With no external information: good protection
• Who is Alice? one of {1,2,3,4,5,6,7,8}
• Alice and Bob connected? 11/28 likelihood 4
2
5
13
6
7
8
12

Adversaries with external information
• Structural knowledge
• often assumed limited to small radius around node
• “Alice has degree 2” or “Bob has two connected neighbors”
• Information can be precise or approximate
• External information may be acquired from a speciﬁc attack, or we may
assume a category of knowledge as a bound on adversary capabilities.
External information: facts about identified individuals
and their relationships in the hidden network.
13

A
B GT G
G
Figure 4 (A) Graph of the largest
component in gang associated STD
outbreak, Colorado Springs,
1989–91 (n = 410). (B) Core of the
largest component in gang
associated STD outbreak, Colorado
Springs, 1989–91 (n = 107).
Sexual network structure i155
Naively Anonymized Network 14
unique or partial
node re-identification
Bob
External information
Alice
Matching attack: the adversary
matches external information to a
naively anonymized network.
Matching attacks

Attacks on naively anonymized networks
• Success of a matching attack depends on:
• descriptiveness of external information
• structural diversity in the network
• With external information: weaker protection
• Who is Alice?
• Who is Alice, if her degree is known to be 4 ?
• Alice and Bob connected?
one of {2,4,7,8}
4
2
5
13
6
7
8
one of {1,2,3,4,5,6,7,8}
15

Local structure is highly identifying
0
0.2
0.4
0.6
0.8
1
H1 H2 H3 H4
Re-identification Risk
FractionofPopulation
Strength of Adversary’s Knowledge
[1]
[2-4]
[5-10]
[11-20]
[>21]
Friendster network
~4.5 million nodes
Uniquely identiﬁed
degree nbrs degrees
Well-protected
[Hay, PVLDB 08] 16

Active attack on an online network
• Goal: disclose edge between two targeted individuals.
• Assumption: adversary can alter the network structure, by creating
nodes and edges, prior to naive anonymization.
• In blogging network: create new blogs and links to other blogs.
• In email network: create new identities, send mail to identities.
• (Harder to carry out this attack in a physical network)
[Backstrom, WWW 07] 17

Active attack on an online network
1
Attacker creates a distinctive
subgraph of nodes and edges.
2
Attacker links subgraph to target
nodes in the network.
Naive anonymizationNaive anonymization
3
Attacker ﬁnds matches for pattern in
naively anonymized network.
4
Attacker re-identiﬁes targets and
discloses structural properties.
Alice
Bob
A
B
G
G
G GT
GT
GT G
GT
G
GTGT
G
GT
G
G
Sexual network structure

Results of active attack
• Given a network G with n nodes, it is possible to construct a
pattern subgraph with k = O(log(n)) nodes that will be unique in G
with high probability.
• injected subgraph is chosen uniformly at random.
• the number of subgraphs of size k that appear in G is small
relative to the number of all possible subgraphs of size k.
• The pattern subgraph can be efﬁciently found in the released
network, and can be linked to as many as O(log2(n)) target nodes.
• In 4.4 million node Livejournal friendship network, attack succeeds
w.h.p. for 7 pattern nodes.

Auxiliary network attack
• Goal: re-identify individuals in a naively anonymized target network
• Assumptions:
• An un-anonymized auxiliary network exists, with overlapping
membership.
• There is a set of seed nodes present in both networks, for which
the mapping between target and auxiliary is known.
• Starting from seeds, mapping is extended greedily.
• Using Twitter (target) and Flickr (auxiliary), true overlap of ~30000
individuals, 150 seeds, 31% re-identiﬁed correctly, 12% incorrectly.
[Narayanan, OAKL 09] 20

Summary
• Naive anonymization may be good for utility...
• ... but it is not sufﬁcient for protecting sensitive information in
networks.
• an individual’s connections in the network can be highly
identifying.
• external information may be available to adversary from outside
sources or from speciﬁc attacks.
• Conclusion: stronger protection mechanisms are required.
21

Questions & challenges
• What is the correct model for adversary external information?
• How do attributes and structural properties combine to increase
identiﬁability and worsen attacks?
• Are there additional attacks on naive anonymization (or other forms
of anonymization)?
Next: How can we strengthen the protection offered
by a released network while preserving utility ?
22

• Understanding privacy risk
• Managing privacy controls
Outline of tutorial
23

Releasing data vs. statistics
A
C
B
J
I
W
D
EF
G
Y
Aa
Bb
M
Dd
Cc
H
P
N
O
Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh
• Releasing transformed networks
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh safe answer
Q
• Releasing “safe” network statistics
To prevent adversary attack,
release transformed network
• transformations obscure
identifying node features
• while hopefully preserve
global topology
24

• A graph G( V, E ) is k-degree anonymous if every node in V has the
same degree as k-1 other nodes in V.
Transform for degree anonymity
25
!"#$%"
&"#'%" ("#'%"
)"#'%" *"#'%"
!"#$%"
&"#$%" ("#$%"
)"#'%" *"#'%"
!"#"$%&'!(#")
[Liu, SIGMOD 08]
1-degree
anonymous
2-degree
anonymous

• Problem: Given a graph G( V, E ) and integer k, ﬁnd minimal set of
edges E’ such that graph G( V, E ∪ E’) is k-degree anonymous.
• Approach: Use dynamic programming to ﬁnds minimum change to
degree sequence.
• Challenge: may not be possible to realize degree sequence through
edge additions.
• Example: V = {a, b, c}, E = { (b,c) }. Degree sequence is [0,1,1].
Min. change yields [1,1,1] but not realizable (without self-loops).
• Algorithm: draws on ideas from graph theory to construct a graph
with minimum, or near minimum, edge insertions.
Algorithm for degree anonymization
26
[Liu, SIGMOD 08]

• Degree anonymization is an instance of a more general paradigm.
Many approaches proposed follow this paradigm.
A common problem formulation
27
Given input graph G,
• Consider set of graphs G, each G* in G reachable from
G by certain graph transformations
• Find G* in G such that G* satisﬁes privacy( G*, ... ), and
• Minimizes distortion( G, G* )

Privacy as resistance to attack
• Adversary capability: knowledge of...
• attributes
• degree
• subgraph neighborhood
• structural knowledge beyond immediate neighborhood
• Attack outcome
• Node re-identiﬁcation
• Edge disclosure
28

Kinds of transformations
• Transformations considered in literature can be classiﬁed into three
categories
• Directed alteration
• Generalization
• Random alteration
29

Directed alteration
• Transform network by adding (or removing) edges
• [Liu, SIGMOD 08] insert edges to achieve degree anonymity
• [Zhou, ICDE 08] neighborhood anonymity, labels on nodes
• [Zou, PVLDB 09] complete anonymity (k isomorphic subgraphs)
• [Cheng, SIGMOD 10] complete anonymity and bounds on edge disclosure
30
Hh
Z
X
Ff
K
S
L
V
Gg
U
Ee
T
R
QO
N
P H
Cc
Dd
M
BbAa
Y
G
F
E
D
W
I
J
B
C
AA
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh

3
4
3
1
3
3
16
31
6
2
8
5
11
9
21
13
24
7
30
32
22
15
26
27
12
10 1 29
25
17
20
14
34
4
19
3
33
28
23
18
1
7
6
3
5
2
1
2
3
1
3
6
4 4
31
5
1
Generalization
• Transform network by cluster nodes into groups
• [Cormode, PVLDB 08] attribute-based attacks (graph structure unmodiﬁed)
on bipartite graphs, prevents edge disclosure
• [Cormode, PVLDB 09] similar to above but for arbitrary interaction graphs
(attributes on nodes and edges)
• [Hay, PVLDB 08, VLDBJ 10] summarize graph topology in terms of node
groups; anonymity against arbitrary structural knowledge
31
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh

Hh
Z
X
Ff
K
S
L
V
Gg
U
Ee
T
R
QO
N
P H
Cc
Dd
M
BbAa
Y
G
F
E
D
W
I
J
B
C
A
Random alteration
• Transform network by stochastically adding, removing, or rewiring edges
• [Ying, SDM 08] random rewiring subject to utility constraint (spectral
properties of graph must be preserved).
• [Liu, SDM 09] randomization to hide sensitive edge weights
• [Wu, SDM 10] exploits spectral properties of graph data to ﬁlter out some of
the introduced noise.
32
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh

Other work in network transformation
• Other works
• [Zheleva, PinKDD 07] predicting sensitive hidden edges from released
graph data (nodes and non-sensitive edges).
• [Ying, SNA-KDD 09] comparison of randomized alteration and directed
alteration.
• [Bhagat, WWW 10] releasing multiple views of a dynamic social network.
• Surveys:
• [Liu, Next Generation Data Mining 08]
• [Zhou, SIGKDD 08]
• [Hay, Privacy-Aware Knowledge Discovery 10]
• [Wu, Managing and Mining Graph Data 10]
33

Evaluating impact on utility
• After transformations, graph is released to public. Analyst
measures transformed graph in place of original. What is impact
on utility?
• Graph remains useful if it is “similar” to original. How measure
similarity?
• Related questions arise in statistical modeling of networks and
assessing model ﬁtness [Goldenberg, Foundations 10] [Hunter, JASA 08]
• Common approach to evaluating utility: empirically compare
transformed graph to original graph in terms of various network
properties
34

Impact on network properties
0 10 20 30
1101001000
Inf 2 4 6 8 10
050100150
0.0 0.2 0.4 0.6 0.8 1.0
01101001000
path lengths clusteringdegree
frequency
!
! !
!
!
!
! ! !
!
0.000.250.50
Avg.Distortion k
2 5 10 20 |V|
0.000.250.50
Clustering
Degree
Paths
Original
k=10
k= |V|
Algorithm from Hay PVLDB 08;
experiments on version of HepTh
network (2.5K nodes, 4.7K edges)
[Hay, PVLDB 08] 35

Limitations
• Utility
• Transformation may distort some properties: some analysts will
find transformed graph useless
• Lack of formal bounds on error: analyst uncertain about utility
• Privacy
• Defined as resistance to a specific class of attacks; vulnerable to
unanticipated attacks?
• Inspired by k-anonymity; doomed to repeat that history? (See
survey [Chen, Foundations and Trends in Database 09].)
36

Outline of tutorial
• Differential privacy
• Degree sequence
• Subgraph counts
37

Releasing data vs. statistics
Ease of use good
Protection anonymity
Accuracy no formal guarantees
A
C
B
J
I
W
D
EF
G
Y
Aa
Bb
M
Dd
Cc
H
P
N
O
Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh
• Releasing transformed networks
A
C
B
J
I
W
D
E
F
G
Y
Aa Bb
M
Dd
Cc
HP
N
O Q
R
T
Ee
U
Gg
V
L
S
K
Ff
X
Z
Hh safe answer
Q Ease of use bad for practical analyses
Protection formal privacy guarantee
Accuracy provable bounds
• Releasing “safe” network statistics
Q(G) + noise
output
perturbation 38

When are aggregate statistics safe to release?
• “Safe” statistics should report on properties of a group, without
revealing properties of individuals.
• We often want to release a combination of statistics. Still safe?
• What if adversary uses external information along with statistics?
Still safe?
• Dwork, McSherry, Nissim, Smith [Dwork, TCC 06] proposed
differential privacy as a rigorous standard for safe release.
• Many existing results for tabular data; relatively few results for
network data.
39

The differential guarantee
Two databases are neighbors if they differ by at most one tuple
D
q
A
DATA OWNER ANALYST
Neighbors
indistinguishable
given output
(no. of ‘B’ students)
name gender grade
Alice Female A
Bob Male B
Carl Male A
q(D)~
D’
q(D’)
q
A
name gender grade
Alice Female A
Carl Male A
~
(noisy answer on D)
40

Pr[A(D) ∈ S] ≤ e
Pr[A(D
) ∈ S]
Differential privacy
A randomized algorithm A provides ε-differential privacy if:
for all neighboring databases D and D’, and
for any set of outputs S:
epsilon is a
privacy parameter
= 0.1 e
≈ 1.10Epsilon is usually small: e.g. if then
epsilon = stronger privacy
[Dwork, TCC 06] 41

Calibrating noise
• How much noise is necessary to ensure differential privacy?
• Noise large enough to hide “contribution” of individual record.
• Contribution measured in terms of query sensitivity.
42

Query sensitivity
where D, D’ are any two neighboring databases
The sensitivity of a query q is
!q = max | q(D) - q(D’) |
D,D’
Query q Sensitivity !q
q1: Count tuples 1
q2: Count(‘B’ students) 1
q3: Count(students with property X) 1
q4: Median(age of students) ~ max age
[Dwork, TCC 06] 43

q(D) + Laplace( scale )Δq / ε
The Laplace mechanism
The following algorithm for answering q is ε-differentially private:
A
Laplace
Mechanism
[Dwork, TCC 06]
true
answer
sample from scaled
distribution
privacy
parameter
sensitivity of q
0
0.25
0.5
-5 -4 -3 -2 -1 0 1 2 3 4 5
0
0.25
0.5
-5 -4 -3 -2 -1 0 1 2 3 4 5
Δq=1
ε=1.0
Bob out Bob inBob out Bob in
Δq=1
ε=0.5
44

Differentially private algorithms
• Any query can be answered (but perhaps with lots of noise)
• Noise determined by privacy parameter epsilon and the sensitivity
(both public)
• Multiple queries can be answered (details omitted)
• Privacy guarantee does not depend on assumptions about the
adversary (caveats omitted, see [Kifer, SIGMOD 11])
Survey paper on differential privacy: [Dwork, CACM 10]
45

Adapting differential privacy for networks
• For networks, what is the right notion of “differential object?”
• Hide individual’s “evidence of participation” [Kifer, SIGMOD 11]
• An edge? A set of k edges? A node (and incident edges)?
• More discussion in [Hay, ICDM 09] [Kifer, SIGMOD 11]
• Choice impacts utility
• Existing work considers only edge, and k-edge, differential privacy.
46
A participant’s sensitive information is not a single edge.

What can we learn accurately?
• What can we learn accurately about a network under edge or k-
edge differential privacy?
• Basic approach:
• Express desired task as one or more queries.
• Check query sensitivity
• if High: not promising, but sometimes representation matters.
• if Low: maybe promising, but may still require work.
47

Outline of tutorial
• Degree sequence
• Subgraph counts
48

The degree sequence can be estimated accurately
• Degree sequence: the list of degrees of each node in a graph.
• A widely studied property of networks.
Alice Bob Carol
Dave Ed
Fred Greg Harry
[1,1,2,2,4,4,4,4]
Inverse
cumulative
distribution
Orkut
crawl
orkut
x
CF(x)
0.00.20.40.60.81.0
s
0 1 4 9 19 49 99
Fraction
Degree
[Hay, PVLDB 10][Hay, ICDM 09]
49

Two basic queries for degrees
Frequency of each degreeFrequency of each degree
cnti count of nodes with degree i
F [cnt0, cnt1, ... cntn-1]
Alice Bob Carol
Dave Ed
Fred Greg Harry
Alice Bob Carol
Dave Ed
Fred Greg Harry
G G’
Degree of each nodeDegree of each node
degA degree of node A
D [degA, degB, ... ]
D(G) = [1,4,1,4,4,2,4,2]
D(Gʼ) = [1,4,1,3,3,2,4,2]
ΔD=2
F(G) = [0,2,2,0,4,0,0,0]
F(Gʼ) = [0,2,2,2,2,0,0,0]
ΔF=4 50

1 2 5 10 20 50 200 5000.00.20.40.60.81.0
orkut
x
Pr[X=x]
These queries are both ﬂawed
• D requires independent samples
from Laplace(2/ε) in each
component.
• F requires independent samples
from Laplace(4/ε) in each
component.
• Thus Mean Squared Error is Θ(n/ε2)
original
D
F
Fraction
Degree
New technique allows
improved error of O(d log3(n)/ε2)
(where d is # of unique degrees)
51

An alternative query for degrees
Degree of each node, rankedDegree of each node, ranked
rnki return the rank ith degree
S [rnk1, rnk2, ... rnkn ]
S(G) = [1,1,2,2,4,4,4,4]
S(Gʼ) = [1,1,2,2,3,3,4,4]
ΔS=2
Degree of each nodeDegree of each node
degA degree of node A
D [degA, degB, ... ]
D(G) = [1,4,1,4,4,2,4,2]
D(Gʼ) = [1,4,1,3,3,2,4,2]
ΔD=2
Alice Bob Carol
Dave Ed
Fred Greg Harry
Alice Bob Carol
Dave Ed
Fred Greg Harry
G G’
52

S(G) = [10, 10, ....10, 10, 14, 18,18,18,18]
• The output of the sorted degree query is not (in general) sorted.
• We derive a new sequence by computing the closest non-
decreasing sequence: i.e. minimizing L2 distance.
0 5 10 15 20 25
05101520
!
!
!
!
!
! !
! !
!
!
! ! !
!
!
!
!
!
!
!
! ! ! !
! !
! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !
!
! ! ! !
!
!
!
S(G) true degree sequence
noisy observations (! = 2)
inferred degree sequence
0 5 10 15 20 25
05101520
!
!
!
!
!
! !
! !
!
!
! ! !
!
!
!
!
!
!
!
! ! ! !
!
!
!
0 5 10 15 20 25
05101520
! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !
!
! ! ! !
!
!
!
!
!
! !
! !
!
!
! ! !
!
!
!
!
!
!
!
! ! ! !
!
!
!
0 5 10 15 20 25
05101520
! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !
!
! ! ! !
!
!
!
Using the sort constraint
= 19th smallest
degree + noise
Degree
Rank
53

Experimental results, continued
powerlaw
α=1.5, n=5M
livejournal
n=5.3M
orkut
n=3.1M
ε=.001
ε=.01
0.00.20.40.60.81.0
0 1 4 9 19 49
0.00.20.40.60.81.0
0 1 4 9 19 49
0.00.20.40.60.81.0
0 1 4 9 19 49 99
0.00.20.40.60.81.0
0 1 4 9 19 49 99
0.00.20.40.60.81.0
0 4 9 49 499
0.00.20.40.60.81.0
0 4 9 49 499
original
noisy
inferredFraction
Degree
54

Outline of tutorial
• Degree sequence
• Subgraph counts
55

Subgraph counting queries
• Given query graph H, return the number of subgraphs of G that are
isomorphic to H.
• Importance
• Used in statistical modeling: exponential random graph models
• Descriptive statistics: clustering coefﬁcient from 2-star, triangle
56
2-star 3-star triangle 2-triangle

Subgraph counts have high sensitivity
• QTRIANGLE: return the number of triangles in the graph
• High sensitivity due “pathological” worst-case graph. If input is not
pathological, can we obtain accurate answers?
...
n-2 nodes
A B
...
n-2 nodes
A B
G G’
QTRIANGLE (G) = 0 QTRIANGLE (G’) = n-2
High Sensitivity:
ΔQTRIANGLE=O(n)
57

Local sensitivity
• Tempting, but ﬂawed, idea is to add noise proportional to local
sensitivity.
• Local sensitivity of q on G: maximum difference in query answer
between G and a neighbor G’.
• Example shows problem of using local sensitivity (from [Smith, IPAM
10]): database D is set of number, query q is the median
LS(G) = max | q(G) - q(G’) |
G’∈N(G)
0...0 000 c...c
}
}
(n-3)/2 (n-3)/2
D =
LS(D)=0
0...0 00c c...cD’ =
}
}
(n-3)/2 (n-3)/2
LS(D’)=c
58

Instance-based noise
• Two general approaches to adding instance-based noise
• Smooth sensitivity Compute a smooth upper bound on local
sensitivity [Nissim, STOC 07].
• Noisy sensitivity Use differentially private mechanism to get
noisy upper “bound” on local sensitivity [Behoora, PVLDB 11]
[Dwork, STOC 09].
• Instance-based noise can require modest relaxation of differential
privacy to account for (very low probability) “bad” events.
59

Differentially private subgraph counts
• For k-stars and triangles, smooth sensitivity is efﬁciently
computable
• For k-triangles with k ! 2
• Computing smooth sensitivity NP-Hard.
• However, it can be estimated using noisy sensitivity approach
• Empirical and theoretical analysis:
• Generally, instance-based noise not much larger than local
sensitivity
• However, for k-triangles on real data, local sensitivity sometimes
large (relative to actual number of k-triangles).
[Behoora, PVLDB 11] 60

Alternative representations
• Number of k-stars in a graph can be computed from the degree
sequence
• In other words, an answer to the high sensitivity k-star query can
be derived from results of the degree sequence estimator.
• Would be interesting to compare error of this approach with
instance-based noise approach of [Behoora, PVLDB 11].
k-stars(G) =

v∈G

deg(v)
k

61

Other work on releasing network statistics
• [Rastogi, PODS 09] Subgraph counting queries under an alternative
model of adversarial privacy. Expected error Θ(log2n) instead of
Θ(n) for restricted class of adversaries.
• [Machanavajjhala, PVLDB 11] Investigates recommender systems that
use friends’ private data to make recommendations.
• Lower bound on accuracy of differentially private recommender
• Experimental analysis shows poor utility under reasonable
privacy.
62

Open questions
• For graph analysis X, how accurately can we estimate X under
edge or node differential privacy?
• Lower bounds on accuracy under node differential privacy?
• Is it socially acceptable to offer weaker privacy protection to high-
degree nodes (as in k-edge differential privacy)?
• Can we generate accurate synthetic networks under differential
privacy?
63

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
– +%,%-#,-*)3@*/#$%6)*63,3(*
– 8@;;%)*%,?*3/',*A@'03,0*
BC*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
•  D=%*#0*/#$%6)*#04*3*3,(#,'*036#%(E,'234#,-*@0'0*
•  F='*0%?*0#@%3,**
– ,?'0%,?#,-*)3@*/#$%6)*#04*
– +%,%-#,-*)3@*/#$%6)*63,3(*
– 8@;;%)*%,?*3/',*A@'03,0*
BG*

-*)+.,/)(01,+%2(%(0)3%,$('#4)+50(
+#((#3,0*3:*@0'0*0=%'*?'%#(0*3:*='#*/'03,%(*(#$'0*
2#=*$%0*,'2340*3:*:#',?0H*%,?*3I',H*0%,-'0*
J3@'0)*3K*=L/KMM222N63,#ON%,?'2N6;@N'?@MPQ.-6M;)-3@/N=;(*
BB*

61,#(%0(7+%8,39(+%05(#)(0)3%,$:'#4)+5%2(0'+0;(
1'(%*)+.,/)(9)(01,+'('=7$%3%#$9('?2?(,.'(,2'(2'@'+(71)'(
,@@+'00('.7$)9'+('#3?(3,($',@(#)(%@'/#9(#1'A?((
BQ*

61,#(%0(7+%8,39(+%05(#)(0)3%,$:'#4)+5%2(0'+0;(
1'(%*)+.,/)(9)(@%@()#(01,+'('=7$%3%#$9(3,(,$0)(B'(@'+%8'@(*+).(9)+(7B$%3(
7+)C$'(*+%'@01%7(3)'3/)0()+('8'(.%3+):#,+2'#'@(,@8'+/0%2(090#'.0?( BR*

1'(0,@(0%#,/)D(
E)4(#)(7+'8'#(.9(
'=(*+).(0''%2(.9(
0#,#0(7@,#'0;((
E)4(#)(1%@'(
.9(*+%'@($%0#(%(
#1'(0',+31(
+'0$#0;(
E)4(#)(7+'8'#(#1'(
,77$%3,/)0(.9(
*+%'@0(%0#,$$'@(*+).(
,33'00%2(.9(
%*)+.,/);(
F$$(.9(*+%'@0(1,8'(
01,+'@(#1'%+(
1).'#)4(,@(71)'(
.B'+(.,9B'(-(
01)$@(,$0)(@)(#1%0;(
-('G)9'@(01,+%2(
.9(@,%$9(,3/8%/'0(
4%#1(#1'(6)+$@H((
I#(,9(,@8'+0'(
'J'3#0;**
K9(L)@H(61,#(
%*)+.,/)(-(1,8'(
01,+'@(,$$(#1'0'(9',+0(
,@(41)(3,(8%'4(
#1'0'(%*)+.,/);(
BS*

1'(0,@(0%#,/)D(M3)#?N(
•  T3@*=%$'*63,3(*3,*2=%*#,:3;%3,*)3@*2%,*3*
0=%'H*2=3*)3@*2%,*3*63,,'6*2#=*
•  T3@*?3*,3*=%$'*63;/'=',0#$'*%,?*%66@%'*#?'%*3:*
='*#,:3;%3,*)3@*=%$'*'U/(#6#()*%,?*#;/(#6#()*
?#06(30'?*
•  8'V,-*3,(#,'*/#$%6)*#0*;'*63,0@;#,-*%,?*;%,)*3:*
)3@*6=330'*3*%66'/*='*?':%@(*0'V,-*
•  .$',@%(()*)3@*(30'*63,3(WN%,?*)3@*%'*:%6#,-*
/#$%6)*#04*
QX*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*'U/(#6#()*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*#;/(#6#()**
•  F33(0*3*$#0@%(#Y'*)3@*/#$%6)*/3(#6#'0**
– +%,%-#,-*)3@*/#$%6)*63,3(*
– 8@;;%)*%,?*3/',*A@'03,0*
QZ*

O+%8,39(+%05(@'(#)(41,#(9)(01,+'@('=7$%3%#$9(
•  !#$%6)*#04*#0*;'%0@'?*O)*!#$%6)*863'*[#@H*9J5+*XS]*
•  !#$%6)*063'*%4'0*#,3*%663@,*2=%*#,:3*)3@^$'*
0=%'?*%,?*2=3*6%,*$#'2*=%*#,:3*
Q_*

I,0%3(7+'.%0'0()*(7+%8,39(03)+'(
•  P'0%/8%#9Q(F='*;3'*0',0#$'*='*#,:3;%3,*
'$'%('?*O)*%*@0'H*='*=#-='*=#0*/#$%6)*#04N*
•  R%0%B%$%#9Q(F='*2#?'*='*#,:3;%3,*%O3@*%*@0'*
0/'%?0H*='*=#-='*=#0*/#$%6)*#04N**
Q`*
mother’s maiden name is more sensitive than mobile-phone number
home address known by everyone poses higher risks than by friends
[#@H*9J5+*XS]*

1'(*+,.'4)+5(
QC*
Privacy Score of User j due to Profile Item i
sensitivity of profile item i visibility of profile item i
name, or gender, birthday, address,
phone number, degree, job, etc.

1'(*+,.'4)+5(M3)#?N(
QG*
Privacy Score of User j due to Profile Item i
sensitivity of profile item i visibility of profile item i
Overall Privacy Score of User j
name, or gender, birthday, address,
phone number, degree, job, etc.

1'(%#'.(+'07)0'(#1')+9(M-SN(,77+),31(
R(n, N)R(n, 1)
R(i, j)
R(1, N)R(1, 2)R(1, 1)
User_1 User_j User_N
Profile Item_1
(birthday)
Profile Item_i
(cell phone #)
Profile Item_n
share, R(i, j) = 1
not share, R(i, j) = 0 Profile item’s discrimination
User’s attitude,
e.g., conservative or extrovert
Profile item’s sensitivity
Profile item i’s true visibility
QB*

T,$3$,/2(7+%8,39(03)+'(0%2(-S(
QQ*
Sensitivity: Visibility:
Overall Privacy Score of User j
byproducts: profile item’s discrimination and user’s attitude
a((*/%%;''0*6%,*O'*'0;%'?*@0#,-**
+%U#;@;*#4'(#=33?*.0;%3,*%,?*.U/'6%3,E+%U#;#Y%3,N*

-#'+'0/2(+'0$#0(*+).(0'+(0#@9(
QR*
Sensitivity of The Profile Items Computed by IRT Model
Survey
Information-sharing preferences of 153
users on 49 profile items such as name,
gender, birthday, political views, address,
phone number, degree, job, etc. are
collected.
Statistics
• 49 profile items
• 153 users from 18 countries/regions
• 53.3% are male and 46.7% are female
• 75.4% are in the age of 23 to 39
• 91.6% hold a college degree or higher
• 76.0% spend 4+ hours online per day
Average Privacy Scores Grouped by Geo Regions

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*'U/(#6#()*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*#;/(#6#()**
•  F33(0*3*$#0@%(#Y'*)3@*/#$%6)*/3(#6#'0**
– +%,%-#,-*)3@*/#$%6)*63,3(*
– 8@;;%)*%,?*3/',*A@'03,0*
QS*

O+%8,39(+%05(@'(#)(41,#(9)(01,+'@(%.7$%3%#$9(
•  !#$%6)*#04*#0*;'%0@'?*O)*=32*;@6=*)3@*/#$%'*
#,:3;%3,*6%,*O'*#,:''?*
•  !#$%'*#,:3;%3,*6%,*O'*#,:''?*:3;**
–  T3@*/@O(#6*/3b('H*:#',?0=#/0H*-3@/*;';O'0=#/0H*'6N**
•  !#$%'*#,:3;%3,*6%,*O'*#,:''?*@0#,-*
–  +%c3#)*$3,-*[d'64'H*D_8!*XS]H*[e='('$%H*DDD*XS]**
–  J3;;@,#)*?''63,*[+#0(3$'H*D85+*ZX]*
–  J(%00#b6%3,*[e='('$%H*DDD*XS]H*[#,?%;33?H*DDD*XS]*
RX*

-*'+'3'(,U,35Q(.,G)+%#9(8)/2(
d%0#6*!';#0'K*O#?0*3:*%*:'%='*f364*3-'='*
RZ*
[d'64'H*D_8!*XS]**
[e='('$%H*DDD*XS]**
FI
V
T
W
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*63,0'$%$'*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK**6334#,-H*%0*
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*('-%(#Y'*0%;'*0'U*;%#%-'*
X-#'+'0#0YK*:%0=#3,*%,?*%//%'(*
XL'@'+YK*g*
XO)$%/3,$(R%'40YK*g*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YQ(G'4'$+9(,@(01)'0(
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*'$')*;'*#*b,?*3@*%*6@'*O3)*
#0*63,0'$%$'*%*(#L('*/%*3:*;'*?#'0*
X-#'+'0#0YK**1Ma*
XL'@'+YK1Ma*
XO)$%/3,$(R%'40YK*1Ma*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK*1Ma*

-*'+'3'(,U,35Q(3)..%#9(@'#'3/)(
R_*
[+#0(3$'H*D85+*ZX]**
0'0*2#=*63;;3,*%L#O@'0*3I',*:3;*?',0'*63;;@,#'0N*
FI
V
T
W
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*63,0'$%$'*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK**6334#,-H*%0*
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*('-%(#Y'*0%;'*0'U*;%#%-'*
X-#'+'0#0YK*:%0=#3,*%,?*%//%'(*
XL'@'+YK*g*
XO)$%/3,$(R%'40YK*g*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YQ(G'4'$+9(,@(01)'0(
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*'$')*;'*#*b,?*3@*%*6@'*O3)*
#0*63,0'$%$'*%*(#L('*/%*3:*;'*?#'0*
X-#'+'0#0YK**1Ma*
XL'@'+YK1Ma*
XO)$%/3,$(R%'40YK*1Ma*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK*1Ma*

-*'+'3'(,U,35Q(3)..%#9(@'#'3/)(
R`*
[+#0(3$'H*D85+*ZX]**
0'0*2#=*63;;3,*%L#O@'0*3I',*:3;*?',0'*63;;@,#'0N*
FI
V
T
W
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*63,0'$%$'*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK**6334#,-H*%0*
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*('-%(#Y'*0%;'*0'U*;%#%-'*
X-#'+'0#0YK*:%0=#3,*%,?*%//%'(*
XL'@'+YK*g*
XO)$%/3,$(R%'40YK*g*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YQ(G'4'$+9(,@(01)'0(
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*'$')*;'*#*b,?*3@*%*6@'*O3)*
#0*63,0'$%$'*%*(#L('*/%*3:*;'*?#'0*
X-#'+'0#0YK**1Ma*
XL'@'+YK1Ma*
XO)$%/3,$(R%'40YK*1Ma*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK*1Ma*

-*'+'3'(,U,35Q(3$,00%C3,/)(
RC*
Z0'+( $'2,$%['(0,.'(0'=(.,++%,2'(( '8'+9(/.'(%(C@()#(,(3#'(B)9(???( '=,0(3)0'+8,/8'0( O)$%/3,$(8%'40(
F( ( ( ]( ;(
I( ]( ( ( $%B'+,$(
T( ( ]( ( $%B'+,$(
V( ( ( ]( 3)0'+8,/8'(
FI
V
T
W
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*63,0'$%$'*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK**6334#,-H*%0*
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*('-%(#Y'*0%;'*0'U*;%#%-'*
X-#'+'0#0YK*:%0=#3,*%,?*%//%'(*
XL'@'+YK*g*
XO)$%/3,$(R%'40YK*g*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YQ(G'4'$+9(,@(01)'0(
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*'$')*;'*#*b,?*3@*%*6@'*O3)*
#0*63,0'$%$'*%*(#L('*/%*3:*;'*?#'0*
X-#'+'0#0YK**1Ma*
XL'@'+YK1Ma*
XO)$%/3,$(R%'40YK*1Ma*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK*1Ma*
[e='('$%H*DDD*XS]**

-*'+'3'(,U,35Q(3$,00%C3,/)(
RG*
FI
V
T
W
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*63,0'$%$'*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK**6334#,-H*%0*
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*('-%(#Y'*0%;'*0'U*;%#%-'*
X-#'+'0#0YK*:%0=#3,*%,?*%//%'(*
XL'@'+YK*g*
XO)$%/3,$(R%'40YK*g*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YQ(G'4'$+9(,@(01)'0(
XL'@'+YK*:';%('*
XO)$%/3,$(R%'40YK*(#O'%(*
XL+)7YK*'$')*;'*#*b,?*3@*%*6@'*O3)*
#0*63,0'$%$'*%*(#L('*/%*3:*;'*?#'0*
X-#'+'0#0YK**1Ma*
XL'@'+YK1Ma*
XO)$%/3,$(R%'40YK*1Ma*
XL+)7YK*'U%0*63,0'$%$'0*
X-#'+'0#0YK*1Ma*
[#,?%;33?H*DDD*XS]*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*'U/(#6#()*
•  !#$%6)*#04*?@'*3*2=%*)3@*0=%'?*#;/(#6#()**
•  F33(0*3*$#0@%(#Y'*)3@*/#$%6)*/3(#6#'0**
– +%,%-#,-*)3@*/#$%6)*63,3(*
– 8@;;%)*%,?*3/',*A@'03,0*
RB*

))$0(#)(8%0,$%['(7+%8,39(7)$%3%'0(
•  h#0@%(#Y%3,0*0#-,#b6%,()*#;/%6*@0'0^*
@,?'0%,?#,-*3:*='#*/#$%6)*0'V,-0*[+%YY#%H*Ji9*ZZ]H*
[#/:3?H*Ji9*ZX]*
RQ*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
– +%,%-#,-*)3@*/#$%6)*63,3(*
•  !#$%6)*;%,%-';',*:3*#,?#$#?@%(0*
•  J3((%O3%$'*/#$%6)*;%,%-';',*:3*0=%'?*63,',0**
– 8@;;%)*%,?*3/',*A@'03,0*
RR*

O+%8,39(.,,2'.'#(*)+(%@%8%@,$0(
•  836#%(*,%$#-%3,*[#@H*9J5+*XS]H*[d'0;'H*87!8*ZX]*
•  !'$',,-*#,:'',6'*%L%640*[#,?%;33?H*DDD*XS]*
•  '%,#,-*/#$%6)*/':'',6'0*2#=*(#;#'?*@0'*#,/@0*
[j%,-H*DDD*ZX]H*[8='=%OH*DDD*ZX]*
RS*

P)3%,$(,8%2,/)((
SX*
P3)+'Q(](^(]_( P3)+'Q(](^(]_(
P3)+'Q(](^(]_(
P3)+'Q(](^(]_(
P3)+'Q(](^(]_(
836#%(*,%$#-%3,*='(/0*@0'0*;%4'*O'L'*/#$%6)*?'6#0#3,0*@0#,-*
63;;@,#)*4,32('?-'*%,?*'U/'0'N**
[#@H*9J5+*XS]*
[d'0;'H*87!8*ZX]*

O+'8'/2(%*'+'3'(,U,350(
SZ*
k';3$'M=#?'*#04)*(#,40H*/3b('0*3*-3@/0*=%*
63,#O@'?*;30*3*='*#,:'',6'*%L%640N*
[#,?%;33?H*DDD*XS]*

`',+%2(7+%8,39(7+'*'+'3'0(
S_*
j#-@'*63@'0)*3*@c@,*j%,-*%,?*l#0',*'j'$'N**
[j%,-H*DDD*ZX]*
[8='=%OH*DDD*ZX]*
'%,#,-*/#$%6)*/':'',6'0*2#=*(#;#'?*@0'*#,/@0*%,?*
%@3;%6%(()*63,b-@'*/#$%6)*0'V,-0*:3*='*@0'N*

1'(*+,.'4)+5(
•  h#'2*/#$%6)*/':'',6'*;3?'(*%0*%*6(%00#b'*
–  h#'2*'%6=*:#',?*%0*%*:'%@'*$'63*
–  !'?#6*6(%00*(%O'(*m%((32*3*?',)n*0=%'*3*,3*0=%'o*
•  l')*5'0#-,*p@'03,0K*
–  i32*3*63,0@6*:'%@'0*:3*'%6=*:#',?g*
–  i32*3*03(#6#*@0'*#,/@0*#,*3?'*3*-'*(%O'('?*?%%g*
S`*

T)0#+3/2(*',#+'0(*)+(',31(*+%'@(
SC*
Age Sex G0 G1 G2 G20 G21 G22 G3
Obama
Fan
Pref. Label
(DOB)
(Alice) 25 F 0 1 0 0 0 0 0 1 allow
(Bob) 18 M 0 0 1 1 0 0 0 0 deny
(Carol) 30 F 1 0 0 0 0 0 0 0 ?
L(
L](
La(
Lb(
La(
La](
Laa(
:#',?0*
j#-@'*63@'0)*3*@c@,*j%,-*%,?*l#0',*'j'$'N**
[j%,-H*DDD*ZX]*
[8='=%OH*DDD*ZX]*
%(03*0''*[q3,'0H*87!8*ZX]*
%(03*0''*[5%,'Y#0H*a98'6*XS]*

P)$%3%/2(0'+(%7#0(
•  a04*@0'*3*(%O'(*0/'6#b6*:#',?0*
–  'N-NH*rD3@(?*)3@*(#4'*3*0=%'*)3@*5%'*3:*d#=*2#=*a(#6'*
a?%;0gs*
•  J=330'*#,:3;%$'*:#',?0*@0#,-*%,*%6$'*('%,#,-*
%//3%6=*
–  ,6'%#,)*0%;/(#,-*
SG*

Z3'+#,%#9(0,.7$%2(
•  8%*2#=*(%O'('?*:#',?0*!*%,?*@,(%O'('?*:#',?0*!#*
•  8%;/(#,-*/36''?0*#,*3@,?0*
–  a04*@0'*3*(%O'(*3,'*:#',?*$*:3;*!#**
–  $*6=30',*O%0'?*3,*@,6'%#,)*'0;%'K*
•  F%#,*d%)'0#%,*6(%00#b'*@0#,-*!*
•  j3*'%6=*$*#,*!#H*'0;%'*%''()*+*%,-./*
•  J=330'*$*#,*!#*=%*;%U#;#Y'0**
*0.1-234.3/*5*6%''()*'(7*%''()*6*%,-./*'(7*%,-./*
•  0'*6%,*A@#*%*%,)*;'*
•  F%#,*/':'',6'*;3?'(*mb,%(*6(%00#b'o*@0#,-*!*
–  0'*3*(%O'(*:#',?0*#,*!#*
SB*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
– +%,%-#,-*)3@*/#$%6)*63,3(*
•  !#$%6)*;%,%-';',*:3*#,?#$#?@%(0*
•  J3((%O3%$'*/#$%6)*;%,%-';',*:3*0=%'?*63,',0**
– 8@;;%)*%,?*3/',*A@'03,0*
SQ*

T)$$,B)+,/8'(7+%8,39(.,,2'.'#(
SR*
!=330*m3*3='*0=%'?*63,',o*@/(3%?'?*3*036#%(*,'234#,-*0#'0*%'*
@0@%(()*63,3(('?*O)*0#,-('*@0'0*2=3*%'*,3*='*%6@%(*3*03('*0%4'=3(?'0N*

T)$$,B)+,/8'(7+%8,39(.,,2'.'#(M3)#?N(
•  F='*J=%((',-'*
–  .%6=*63E32,'*;#-=*=%$'*%*?#t'',*%,?*/300#O()*63,%0,-*
/#$%6)*/':'',6'*
–  i32*3*6=330'*/#$%6)*0'V,-*3*;%U#;#Y'*3$'%((*O','b0g*
•  a,*%L';/K*6(%4'*%U*;'6=%,#0;*[8A@#66#%#,#H*DDD*XS]*
–  '%6=*32,'*#,?#6%'0*='*/'6'#$'?*O','b*%*'%6=*/#$%6)*('$'(*
m892-*)439*.(*(.-+*892-*)439*$24-.,8+*-31:o*
–  ='*0)0';*b,?0*='*O'0*/#$%6)*/':'',6'*=%*;%U#;#Y'0*='*
3$'%((*036#%(*O','b*
–  '%6=*32,'*/%)0*6'%#,*%U*3*='*0)0';*3*63;/',0%'*
3='0^*(30'*
–  ='*;'6=%,#0;*/'$',0*%,*32,'*:3;*@,@=:@(()*?'6(%#,-*
='*O','b*3*;%,#/@(%'*3@63;'0*%*='*%?$%,%-'*
SS*
[8A@#66#%#,#H*DDD*XS]*

!#$%'()*(##)+%,$(
•  !#$%'()*+%,%-#,-*.,'/#0'*1'234*5%%*
•  !'03,%(*!#$%6)*#,*7,(#,'*836#%(*1'2340*
– 9,:3;%3,*0=%#,-*#,*036#%(*,'2340*
– ,?'0%,?#,-*)3@*/#$%6)*#04*
– +%,%-#,-*)3@*/#$%6)*63,3(*
•  !#$%6)*;%,%-';',*:3*#,?#$#?@%(0*
•  J3((%O3%$'*/#$%6)*;%,%-';',*:3*0=%'?*63,',0**
– 8@;;%)*%,?*3/',*A@'03,0*
ZXX*

P..,+9(
•  T3@*=%$'*6'%#,*63,3(*3:*='*#,:3*)3@*%'*0=%#,-*
•  T3@*3I',*6%,,3*'0;%'*='*(3,-*';*#04*$0N*
0=3*';*-%#,*
•  a(-3#=;0*3*;'%0@'*/3',%(*/#$%6)*#040*?@'*3*
#,:3*0=%'?*'#='*'U/(#6#()*3*#;/(#6#()*
•  +3?'(0*3*%(('$#%'*)3@*O@?',*3,*/#$%6)*
;%,%-';',*
ZXZ*

!7'(c'0/)0(
•  a*2#?'()*%66'/'?*/#$%6)*063'*=%*O3300*/@O(#6*
%2%','00*3:*='*/#$%6)*#04*
•  a,*',?E3E',?*7+,3/3,$*0)0';*3*;'%0@'*%,?*
;%,%-'*/#$%6)*3,(#,'*
ZX_*

Privately managing
enterprise network data
Personal Privacy in
Online Social Networks
Data: Enterprise collects
data or observes interactions
of individuals.
Control: Enterprise controls
dissemination of information.
Goal: permit analysis of
aggregate properties; protect
facts about individuals.
Challenges: privacy for
networked data, complex
utility goals.
Data: Individuals contribute
their data thru participation
in OSN.
Control: Individuals control
their connections,
interactions, visibility.
Goal: reliable and
transparent sharing of
information.
Challenges: system
complexity, leaks thru
inference, unskilled users.

Open questions and future directions
• Anonymity: models of adversary knowledge, new attacks, new
network transformations, improved utility evaluation.
• Differential privacy: adapting privacy deﬁnition to networks,
mechanisms for accurate estimates of new network statistics,
synthetic network generation, error-optimal mechanisms,
• Extended data model: attributes on nodes/edges, dynamic network
data.

References
• [Backstrom, WWW 07] L. Backstrom, C. Dwork, and J. Kleinberg. Wherefore art
thou r3579x?: anonymized social networks, hidden patterns, and structural
steganography. In WWW 2007.
• [Becker, W2SP 09] J. Becker and H. Chen. Measuring privacy risk in online social
networks. In W2SP 2009.
• [Behoora, PVLDB 11] I. Behoora, V. Karwa, S. Raskhodnikova, A. Smith, G.
Yaroslavtsev. Private Analysis of Graph Structure. In PVLDB 2011.
• [Besmer, CHI 10] A. Besmer and H. Lipford. Moving beyond untagging: photo
privacy in a tagged world. In CHI 2010.
• [Besmer, SOUPS 10] A. Besmer, J. Watson, and H. Lipford. The impact of social
navigation on privacy policy conﬁguration. In SOUPS 2010.
• [Bhagat, WWW 10] S. Bhagat, G. Cormode, B. Krishnamurthy, D. Srivastava.
Privacy in dynamic social networks. In WWW 2010.
• [Campan, PinKDD 08] A. Campan and T. M. Truta. A clustering approach for data
and structural anonymity in social networks. In PinKDD 2008.

References (continued)
• [Chen, Foundations and Trends in Database 09] B. Chen, D. Kifer, K. LeFevre,
and A. Machanavajjhala. Privacy-Preserving Data Publishing. In Foundations and
Trends in Databases 2009.
• [Cheng, SIGMOD 10] J. Cheng, A. Wai-Chee Fu, and J. Liu. K-Isomorphism:
Privacy Preserving Network Publication against Structural Attacks. In SIGMOD
2010.
• [Cormode, PVLDB 08] G. Cormode, D. Srivastava, T. Yu, and Q. Zhang:
Anonymizing bipartite graph data using safe groupings. In PVLDB 2008.
• [Cormode, PVLDB 09] G. Cormode and D. Srivastava and S. Bhagat and B.
Krishnamurthy. Class-based graph anonymization for social network data. In
PVLDB 2009.
• [Danezis, AISec 09] G. Danezis. Inferring privacy policies for social networking
services. In AISec 2009.
• [Dwork, CACM 10] C. Dwork. A ﬁrm foundation for privacy. In CACM 2010.
• [Dwork, STOC 09] C. Dwork and J. Lei. Differential privacy and robust statistics.
In STOC 2009.

• [Dwork, TCC 06] C. Dwork, F. McSherry, K. Nissim, and A. Smith. Calibrating
noise to sensitivity in private data analysis. In TCC 2006.
• [Fang, WWW 10] L. Fang and K. LeFevre. Privacy wizards for social networking
sites. In WWW 2010.
• [Goldenberg, Foundations 10] A. Goldenberg, S. Fienberg, A. Zheng, E. Airoldi. A
Survey of Statistical Network Models. In Foundations 2009.
• [Hay, ICDM 09] M. Hay, C. Li, G. Miklau, and D. Jensen. Accurate estimation of
the degree distribution of private networks. In ICDM 2009.
• [Hay, Privacy-Aware Knowledge Discovery 10] M. Hay and G. Miklau and D.
Jensen. Enabling Accurate Analysis of Private Network Data. In Privacy-Aware
Knowledge Discovery: Novel Applications and New Techniques 2010.
• [Hay, PVLDB 08] M. Hay, G. Miklau, D. Jensen, and D. Towsley. Resisting
structural identiﬁcation in anonymized social networks. In PVLDB 2008.
• [Hay, PVLDB 10] M. Hay, V. Rastogi, G. Miklau, and D. Suciu. Boosting the
accuracy of differentially-private queries through consistency. In PVLDB 2010.

• [Hay, VLDBJ 10] M. Hay and G. Miklau and D. Jensen and D. Towsley and C. Li.
In VLDB Journal 2010.
• [Hunter, JASA 08] D. Hunter, S. Goodreau, and M. Handcock. Goodness of ﬁt of
social network models. In JASA 2008.
• [Jones, SOUPS 10] S. Jones, E. O’Neill. Feasibility of structural network
clustering for group-based privacy control in social networks. In SOUPS 2010.
• [Kifer, SIGMOD 11] D. Kifer and A. Machanavajjhala. No Free Lunch in Data
Privacy. In SIGMOD 2011.
• [Krishnamurthy, WWW 09] B. Krishnamurthy, C. Wills. Privacy diffusion on the
web: a longitudinal perspective. In WWW 2009.
• [Lindamood, WWW 09] J. Lindamood, R. Heatherly, M. Kantarcioglu, B.
Thuraisingham. Inferring private information using social network data. In WWW
2009.
• [Lipford, CHI 10] H. Lipford, J. Watson, M. Whitney, K. Froiland, R. Reeder. Visual
vs. compact: a comparison of privacy policy interfaces. In CHI 2010.

• [Liu, ICDM 09] K. Liu and E. Terzi. A framework for computing privacy scores of
users in online social networks. In ICDM 2009.
• [Liu, Next Generation Data Mining 08] K. Liu, K. Das, T. Grandison, and H.
Kargupta. Privacy-Preserving Data Analysis on Graphs and Social Networks. In
Next Generation of Data Mining 2008.
• [Liu, SDM 09] L. Liu and J. Wang and J. Liu and J. Zhang. Privacy Preservation in
Social Networks with Sensitive Edge Weights. In SDM 2009.
• [Liu, SIGMOD 08] K. Liu and E. Terzi. Towards identity anonymization on graphs.
In SIGMOD 2008.
• [Machanavajjhala, PVLDB 11] A. Machanavajjhala, A. Korolova, and A. Das
Sarma. Personalized Social Recommendations -- Accurate or Private? In VLDB
2011
• [Mazzia, CHI 11] A. Mazzia, K. LeFevre, and E. Adar. A tool for privacy
comprehension. In CHI 2011.

• [Mislove, WSDM 10] A. Mislove, B. Viswanath, K. Gummadi, P. Druschel. You are
who you know: Inferring user proﬁles in online social networks. In WSDM 2010.
• [Narayanan, OAKL 09] A. Narayanan and V. Shmatikov. De-anonymizing social
networks. In Security and Privacy 2009.
• [Nissim, STOC 07] K. Nissim, S. Raskhodnikova, and A. Smith. Smooth
sensitivity and sampling in private data analysis. In STOC 2007.
• [Rastogi, PODS 09] V. Rastogi, M. Hay, G. Miklau, and D. Suciu. Relationship
privacy: Output perturbation for queries with joins. In PODS 2009.
• [Seigneur, Trust Management 04] J. Seigneur and C. Damsgaard Jensen.
Trading privacy for trust, Trust Management 2004
• [Shehab, WWW 10] M. Shehab, G. Cheek, H. Touati, A. Squicciarini, and P.
Cheng. Learning based access control in online social networks. In WWW 2010.
• [Smith, IPAM 10] A. Smith. In IPAM Workshop on Statistical and Learning-
Theoretic Challenges in Data Privacy 2010.

• [Squicciarini, WWW 09] A. Squicciarini, M. Shehab, F. Paci. Collective privacy
management in social networks. In WWW 2009.
• [Wu, Managing and Mining Graph Data 10] X. Wu, X. Ying, K. Liu, and L. Chen. A
Survey of Algorithms for Privacy- Preservation of Graphs and Social Networks. In
Managing and Mining Graph Data 2010.
• [Wu, SDM 10] L. Wu and X. Ying and X. Wu. Reconstruction from Randomized
Graph via Low Rank Approximation. In SDM 2010.
• [Ying, SDM 08] X. Ying and X. Wu. Randomizing social networks: a spectrum
preserving approach. In SDM 2008.
• [Ying, SNA-KDD 09] X. Ying and K. Pan and X. Wu and L. Guo. Comparisons of
Randomization and K-degree Anonymization Schemes for Privacy Preserving
Social Network Publishing. In PinKDD 2009.
• [Zheleva, PinKDD 07] E. Zheleva and L. Getoor. Preserving the privacy of
sensitive relationships in graph data. In PinKDD 2007.

• [Zheleva, WWW 09] E. Zheleva and L. Getoor. To join or not to join: the illusion of
privacy in social networks with mixed public and private user proﬁles. In WWW
2009.
• [Zhou, ICDE 08] B. Zhou and J. Pei. Preserving privacy in social networks against
neighborhood attacks. In ICDE 2009.
• [Zhou, KIS 10] B. Zhou and J. Pei. k-Anonymity and l-Diversity Approaches for
Privacy Preservation in Social Networks against Neighborhood Attacks. In
Knowledge and Information Systems: An International Journal 2010.
• [Zhou, SIGKDD 08] B. Zhou and J. Pei and W. Luk. A Brief Survey on
Anonymization Techniques for Privacy Preserving Publishing of Social Network
Data. In SIGKDD 2008.
• [Zou, PVLDB 09] L. Zou, L. Chen, and M. T. A. Ozsu. K-automorphism: A general
framework for privacy preserving network publication. In PVLDB 2009.

Privacy-Aware Data Management in Information Networks - SIGMOD 2011 Tutorial

More Related Content

Viewers also liked

Similar to Privacy-Aware Data Management in Information Networks - SIGMOD 2011 Tutorial

Recently uploaded

Privacy-Aware Data Management in Information Networks - SIGMOD 2011 Tutorial