personalized privacy protection in social networks

Graph Data Management Lab, School of Computer Science

GDM@FUDAN

Personalized Privacy Protection in Social Networks (VLDB2011)

http://gdm.fudan.edu.cn

徐波 Email : [email protected] 6:00PM1

Speaker : XuBo

Personalized Privacy Protection in Social Networks

2

2011-12-05 6:00PM


GDM@FUDAN


徐波 Email : [email protected]


GDM@FUDAN


Privacy Protection

3

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Privacy Protection in Social Networks more and more people join multiple social

networks on the Web as a service provider, it is essential to protect

users’ privacy and at the same time provide “useful” data

4

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Previous Works Clustering-based approaches Graph editing methods Drawback

• Different users may have different privacy preferences• Same level may not be fair and data may useless

5

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Personalized Privacy Protection in Social Networks define different levels of protection for users and

incorporate them into the same published social network

A user can have a clear estimation of the knowledge that an attacker can know about him

The knowledge an attacker uses to find the privacy information of a user is called the background knowledge

6

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


background knowledge

7

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


A simple example protection objectives

• the probability that an attacker finds a person P is node u in the published graph should be less than 50%

• the probability that an attacker finds person P1 and P2 have a connection is less than 50%

8

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Personalized Protection

9

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Method For Level 1 protection, we use node label

generalization For Level 2 protection, we combine the noise

node/edge adding methods based on the protection at Level 1

For Level 3 protection, we further use the edge label generalization to achieve the protection objectives

10

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Abstract PROBLEM DEFINITION LEVEL 1 PROTECTION LEVEL 2 PROTECTION LEVEL 3 PROTECTION EXPERIMENTS

11

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


PROBLEM DEFINITION G(V,E), Given a constant k

12

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


NP-hard problem

Here the cost stands for the sum of label generalization levels. For example, in Figure 1(d), if let the labels in the group {a,d} to be the same, they become [(Africa, American), 2*], thus the node label generalization cost of this group is 4.

13

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


LEVEL 1 PROTECTION Level 1’s background knowledge is about node

label list generalization is to group nodes and make the

nodes within each group have one generalized node label list

14

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Achieve Level 1 protection if each group’s size is at least k, the first

objective is achieved. To satisfy objectives (2) and (3), the following

condition must be satisfied:

The first condition constrains that any two nodes in the same group do not connect to a same node The second condition constrains that no edges within a group

15

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


The Safety Grouping Condition (SGC)

16

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Algorithm 1: Generate safe groups

17

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


LEVEL 2 PROTECTION Adding nodes/edges into GL1 Label Assignment

18

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Construct a k-degree anonymous graph The first step is to set a target

degree for each node the second step is to randomly

create edges between the nodes which need to increase their degrees under SGC

we add noise nodes to enable all the nodes in GL1 have their target degrees by connecting them with noise nodes under SGC.

we need to hide the noise nodes by making their degree to a preselected value degree target

Degree target is the degree that the maximum number of groups of size k in GL1 have

19

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Algorithm 2: Degree Anonymizing Algorithm with X

20

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


A running example of Algorithm 2

21

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Label Assignment Nocon(A1,A2, l) a node/edge label assignment with less

Nocon(A1,A2, l) changes is preferred heuristic method• first assign labels to the noise edges• decide the labels of noise nodes

22

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Assign labels to the noise edges

23

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Decide the labels of noise nodes

24

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


LEVEL 3 PROTECTION degree label sequence a group contains at least one node that needs Level

3 protection, for any nodes in it, change their degree label sequence to be the same by generalizing edge labels.

25

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Generate degree label anonymous graph

26

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


EXPERIMENTS Utilities

• One hop query (1 hop)• (l1, le, l2)

• Two hop query (2 hop)• (l1, le, l2, le2, l3)

• Relative error = • Degree Distribution• EMD (Earth Mover Distance)

• a good measure of the difference between two distributions

27

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Datasets Speed Dating Data

• 552 nodes and 4194 edges• Each node has 19 labels• Edge label is “match”/”unmatch”

ArXiv Data• 19835 nodes and 40221 edges• Each node denotes an author• set label ”seldom” to edges with weight 1, ”normal” to

edges with weight 2 - 5 and ”frequent” to edges with weight larger than 5

ArXiv Data with uniform labels

28

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Results

higher utility average relative error increases with the

increasing of k

29

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Results(2)

30

2011-12-05 6:00PM


GDM@FUDAN




GDM@FUDAN


Results(3)

personalized privacy protection in social networks

Documents