幻灯片 1 - Peking University

Download Report

Transcript 幻灯片 1 - Peking University

VoteTrust
Leveraging Friend Invitation Graph to Defend
Social Network Sybils
Jilong Xue , Zhi Yang , Xiaoyong Yang, Xiao Wang,
Lijiang Chen and Yafei Dai
Computer Science Department, Peking University
Sybil attack in Social networks
Sybils
Friend invitation
Non-popular users
reject
accept
VoteTrust: An Overview
• Basic idea:
– Considering invitation feedback as voting
• Key techniques:
– Trust-based votes assignment
– Global vote aggregation
• Properties:
– High precision in Sybil detection
– Efficient in limiting Sybil’s attack ability
Graph Model
A
A
1
C
C
B
Link initiation graph 𝐺𝐼 = (𝑉𝐼 , 𝐸𝐼 )
B
0
Link acceptance graph 𝐺𝐸 = (𝑉𝐸 , 𝐸𝐸 , 𝑊𝐸 )
Framework of VoteTrust
• Select trust seed – high reliable users
• Distribute votes
• Collect votes and computing score
Outline
Preliminary
Implementation
– Trust-based vote assignment
– Global vote aggregation
Evaluation
Conclusion
Votes Assignment
• Problem:
– How to distribute votes
across users?
vv
vv
v
Reliable user
• Principle:
– Reliable user should get
more votes
• How to implement?
Non-popular user
Sybil
Trust-based Votes Assignment
• Step1: Assigning votes to little human-selected
reliable seeds
• Step2: Propagating to whole users across the
Link initiation graph
𝒗𝒐𝒕𝒆(𝑢) = 𝑑 ∙
𝑣:(𝑣,𝑢)∈𝐸𝐼
𝒗𝒐𝒕𝒆(𝑣)
+ 1 − 𝑑 ∙ 𝒊𝒏𝒊𝒕(𝑢)
𝑜𝑢𝑡_𝑑𝑒𝑔𝑟𝑒𝑒(𝑣)
Example
A
C
B
D
E
B
C
D
E
t=0
5
0
0
0
0
t=1
0.75
4.25
0
0
0
t=2
0.75
0.65
0
1.80
1.80
t=3
2.59
0.94
0.31
0.58
0.58
t=n
1.69
1.57
0.14
0.80
0.80
…
A
 Node A is reliable seed
 Total votes =5
Outline
Preliminary
Implementation
– Trust-based vote assignment
– Global vote aggregation
Evaluation
Conclusion
Vote Aggregating
• Problem:
– How to collect votes and
compute user trust score?
– Trust score 𝑝 𝑢 ∈ [0,1]
vote=1,score=0.2 vote=1,score=0.9
A
B
0
• Principle:
– Trust user should have
high weight in voting.
1
C
score=?
1 × 0.9
1 × 0.2 + 1 × 0.9
= 0.82
𝑠𝑐𝑜𝑟𝑒(𝐶) =
Global Vote Aggregation
• Step1: Set all users’ initial score as 0.5;
• Step2: Iteratively computing each user’s trust
score according to aggregated votes.
𝒔𝒄𝒐𝒓𝒆 𝑢 =
𝑣𝑜𝑡𝑒 𝑣 ∙ 𝒔𝒄𝒐𝒓𝒆 𝑣 ∙ 𝑥𝑣,𝑢
,
𝑣𝑜𝑡𝑒 𝑣 ∙ 𝒔𝒄𝒐𝒓𝒆 𝑣
(𝑣, 𝑢) ∈ 𝐸𝐸
Small-sample Problem
• Number of votes is
too small.
• Wilson score
1
𝒑+
𝑧1−𝛼 2
2𝑁
𝒑=
1
1 + 𝑧1−𝛼 2
𝑁
– weighted average of
1
𝒑 and .
2
vote=1,score=0.2
A
0
B
score=0 ?
score=0.40
vote=1,score=0.2
A
1
B
score=1 ?
score=0.61
Security Properties (I)
• Theorem 1: The number of Sybil’s attack-link
needs to satisfy the following upper bound
𝑵𝒐𝒖𝒕
𝛿𝑓 − 𝛿𝑓 2
≤ 𝜌𝑵𝒊𝒏 ∙
𝛿𝑓 − 𝑟
where 𝛿𝑓 is detection threshold.
𝑁𝑖𝑛
Sybil
Normal user
𝑁𝑜𝑢𝑡
Simulation of Theorem 1
• Comm size: 100
• # of in-links: 10
• Nout avg: 2.36
• Nout max:4
Security Properties (II)
• Theorem 2: Sybil community size need to
satisfy the upper bound ,
𝑵𝒔 ≤ 𝜎 ∙
𝑵𝒊𝒏
𝛿𝑣
where 𝛿𝑣 is vote collection threshold.
Simulation of Theorem 2
Outline
Preliminary
Implementation
– Trust-based vote assignment
– Global vote aggregation
Evaluation
Conclusion
Experimental Setup
• Data Set
– Renren regional network (PKU) include 200K
users, 5.01 million friend invitations
– 2502 Sybil accounts detected by Renren
– Manual checking 73 Sybils from 500 random user
• Methodology
– Compared with TrustRank and BadRank
– Evaluation metrics: Precision and Recall
TrustRank vs. VoteTrust
Averagely improve 32.9%
Averagely improve 75.6%
BadRank vs. VoteTrust
Averagely improve 44.5%
Averagely improve 41.6%
Separating Normal User from Sybils
80% with low score
Separating Normal User from Sybils
Maximum accuracy=85.7%
Performance Summary
 Outperforms TrustRank by 32.9% in detection
precision averagely;
 Outperforms BadRank by 44.5% in detection
precision averagely;
 High accurate in classifying the Sybil and
normal user (include non-popular user)
Outline
Preliminary
Implementation
– Trust-based vote assignment
– Global vote aggregation
Evaluation
Conclusion
Conclusion
• VoteTrust is a rating system
– high accuracy in Sybil detection
– Efficient in resisting Sybil (community)
• Key techniques
– Trust-based vote assignment
– Global vote aggregation
Thank you!