private analysis of graph structure
DESCRIPTION
Private Analysis of Graph Structure. Grigory Yaroslavtsev http://grigory.us. With Vishesh Karwa , Sofya Raskhodnikova and Adam Smith Pennsylvania State University. Publishing network data. Many data sets can be represented as a graph : Friendship in online social network - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/1.jpg)
1
Private Analysis of Graph Structure
With Vishesh Karwa, Sofya Raskhodnikova and Adam Smith
Pennsylvania State University
Grigory Yaroslavtsev http://grigory.us
![Page 2: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/2.jpg)
2
• Publish information about a graph
• Preserve privacy of relationships
Publishing network dataMany data sets can be represented as a graph:• Friendship in online social network• Financial transactions • Romantic relationships
American J. Sociology, Bearman, Moody, Stovel
Naïve approach: anonymization
![Page 3: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/3.jpg)
3
Goal: Publish structural information about a graph
Publishing network data
DatabaseRelationships Users
Aqueries
answers)(
Government,researchers,businesses
(or) Maliciousadversary
• Anonymization not sufficient [Backström, Dwork, Kleinberg ’07, Narayanan, Shmatikov ’09, Narayanan, Shi, Rubinstein ’11]• Ideal: Algorithms with rigorous privacy guarantee, no assumptions about attacker’s prior information/algorithm
![Page 4: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/4.jpg)
4
• Limits incremental information by hiding presence/absence of an individual relationship
DatabaseRelationships Users
Aqueries
answers)(
Government,researchers,businesses
(or) Maliciousadversary
Differential privacy [Dwork, McSherry, Nissim, Smith ’06]
• Neighbors: Graphs G and G’ that differ in one edge• Answers on neighboring graphs should be similar
![Page 5: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/5.jpg)
5
-differential privacy (edge privacy)For all pairs of neighbors and all events S:
Differential privacy for relationships
𝑃𝑟 [ 𝐴 (𝐺 )∈𝑺 ]≤𝑒𝜖Pr [𝐴 (𝐺′ )∈𝑺 ]
A(G) A(G’)
• Probability is over the randomness of A• Definition requires that the distributions are close:
![Page 6: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/6.jpg)
6
For graphs G and H: # of occurrences of H in GSubgraph counts
Example:
Total: 40
Total: 2
Total: 1
Triangle:
2-star:
2-triangle:
k-star
…k
k-triangle
k…
![Page 7: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/7.jpg)
7
• Subgraph counts are used in:– Exponential random graph models– Descriptive graph statistics, e.g.:
Clustering coefficient =
• Our focus: efficient differentially private algorithms for releasing subgraph counts
Subgraph counts
##
![Page 8: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/8.jpg)
8
Previous work• Smooth Sensitivity [Nissim, Raskhodnikova, Smith ‘07]– Differentially private algorithm for triangles– Open: private algorithms for other subgraphs?
• Private queries with joins [Rastogi, Hay, Miklau, Suciu ‘09]– Works for a wide range of subgraphs– Weaker privacy guarantee, applies only for specific
class of adversaries• Private degree sequence [Hay, Li, Miklau, Jensen ’09]– Guarantees differential privacy– Works for k-stars, but not for other subgraphs
![Page 9: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/9.jpg)
9
Laplace Mechanism and Sensitivity[Dwork, McSherry, Nissim, Smith ‘06]
• Add noise: mean = 0, standard deviation , where is sensitivity => -differential privacy:
• Local sensitivity ([NRS’07], not differentially private!):
• Previous work (mostly): Global sensitivity
𝑓 ′ (𝐺 )= 𝑓 (𝐺 )+𝐿𝑎𝑝 (𝑺 𝒇 /𝝐)
𝐿𝑆𝑓 (𝐺 )= max𝐺′ :𝐍𝐞𝐢𝐠𝐡𝐛𝐨𝐫 𝑜𝑓 𝐺
|𝑓 (𝐺 )− 𝑓 (𝐺′ )|
differentially private!
![Page 10: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/10.jpg)
10
𝑆 𝑓∗( G )
𝐿𝑆 𝑓 ( G )
Instance-Specific Noise = set of all graphs on n vertices. d(G,G’) = # edges in which G and G’ differ.
• Add Cauchy noise: median = 0, median absolute value (where ) => -differential privacy:
• Naïve computation requires exponential time• [NRS’07]: Compute smooth sensitivity for triangles
Smooth Sensitivity [Nissim, Raskhodnikova, Smith ’07]: =
𝑓 ′ (𝐺 )= 𝑓 (𝐺 )+ h𝐶𝑎𝑢𝑐 𝑦 (𝑺 𝒇 ,𝜷∗ / 𝜷)
![Page 11: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/11.jpg)
11
Our contributions
• Differentially private algorithms for k-stars and k-triangles– Efficiently compute smooth sensitivity for k-stars– NP-hardness for k-triangles and k-cycles– Different approach for k-triangles
• Average-case analysis in G(n,p) • Theoretical comparison with previous work• Experimental evaluation
![Page 12: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/12.jpg)
12
Smooth Sensitivity for k-stars ( )
This paper: near-linear time algorithm for smooth sensitivity• Algorithm also reveals structural results, e.g.:– Proposition:
If ) and (maximum degree > then (smooth sensitivity) = (local sensitivity)
• Algorithm optimal for large class of graphs– Proposition: error > (local sensitivity)
• Compared to [HLMJ’09] (private degree sequence):– Our error never worse by more than a constant factor– For 2-stars, our error can be better by factor
…
![Page 13: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/13.jpg)
13
Private Approximation to Local Sensitivity: k-triangles ( )
Approximate differential privacy, -privacy [Dwork, Kenthapadi, McSherry, Mironov, Naor ‘06]:
Pr [ 𝐴 (𝐺 )∈𝑺 ]≤𝑒𝜖Pr [𝐴 (𝐺′ )∈𝑺 ]+𝛿
Idea: Private upper bound on local sensitivity ().Release: , ).
…
If• is -differentially private and • Then A is ()-differentially private.
![Page 14: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/14.jpg)
14
Evaluating our algorithms
• Theoretical evaluation in G(n,p) model– All of our algorithms have relative error -> 0
when the average degree = grows• Empirical evaluation– Synthetic graphs from G(n,p) model– Variety of real data sets
![Page 15: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/15.jpg)
Experimental results for G(n,p)• Comparison with previous work for
0 100 200 300 400 500 600 700 800 90010000.001
0.01
0.1
1
10 2-StarsLS Barrier
Our algorithms
HLMJ
RHMS Lower
RHMS upper
Relative Error = 1
5 % Relative Error
Nodes
Rela
tive
Med
ian
Erro
r
![Page 16: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/16.jpg)
Experimental results for G(n,p)• Comparison with previous work for
0 100 200 300 400 500 600 700 800 900 10000.1
13-Stars
LS Barrier
Our algorithms
HLMJ
RHMS Lower
RHMS upper
Relative Error = 1
20 % Relative Error
Nodes
Rela
tive
Med
ian
Erro
r
![Page 17: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/17.jpg)
Experimental results for G(n,p)
0 200 400 600 800 1000
10
2 Triangles
0 200 400 600 800 10001
100 Triangles
Rela
tive
Med
ian
Erro
r
01000
10
LS Barrier Our algorithms RHMS Lower
RHMS upper Relative Error = 1
• Comparison with [RHMS’09] for
![Page 18: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/18.jpg)
Experimental results (SNAP)
ca-G
rQc
ca-H
epTh
ca-C
ondM
at
ca-H
epPh
Emai
l-Enr
on
ca-A
stro
Ph
n=5K n=10K n=23K n=12K n=37K n=19Km=29K m=52K m=187K m=237K m=368K m=396K
0.0001
0.001
0.01
0.1
1
102-triangles triangles
Rela
tive
Med
ian
Erro
r
![Page 19: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/19.jpg)
Experimental results (SNAP)
ca-G
rQc
ca-H
epTh
ca-C
ondM
at
ca-H
epPh
Emai
l-Enr
on
ca-A
stro
Ph
n=5K n=10K n=23K n=12K n=37K n=19Km=29K m=52K m=187K m=237K m=368K m=396K
0.0001
0.001
0.01
0.1 2-stars HLMJ - 2-stars
Rela
tive
Med
ian
Erro
r
![Page 20: Private Analysis of Graph Structure](https://reader036.vdocument.in/reader036/viewer/2022081501/568139d2550346895da183f2/html5/thumbnails/20.jpg)
20
Summary
• Private algorithms for subgraph counts– Rigorous privacy guarantee (differential privacy)– Running time close to best algorithms for computing
the subgraph counts• Improvement in accuracy and (for some graph
counts) in privacy over previous work• Techniques:– Fast computation of smooth sensitivity– Differentially private upper bound on local sensitivity