The document discusses several algorithms for reputation and ranking systems on graphs, including blogs and social networks. It begins with discussing the Sybil attack and SybilGuard for detecting such attacks. It then covers algorithms for ranking blogs like BlogRank, B2Rank, and EigenRumor that take into account factors like blog activity, links, and comments. MailRank and TrustRank are discussed for using reputation propagation and spam scores to detect spam. The challenges of building measurable and dynamic reputation systems that avoid an arms race are also summarized.
Report
Share
Report
Share
1 of 44
Download to read offline
More Related Content
Reputation Systems II
1. Reputation Systems II
Sybil Attack, BlogRank, B2Rank, EigenRumor,
MailRank, TrustRunk
Yury Lifshits
Caltech
http://yury.name
Caltech CMI Seminar
March 4, 2008
1 / 22
4. Sybil Attack
Graph of trust-weighted edges
n honest nodes + adversary
overall trust value on attack edges
(honest-malicious) is limited
4 / 22
5. Sybil Attack
Graph of trust-weighted edges
n honest nodes + adversary
overall trust value on attack edges
(honest-malicious) is limited
Question: whether splitting adversarial node
into many is beneficial for acquiring higher
reputation (rank)?
4 / 22
6. Negative Result
Assume reputation scores remain the same
under isomorphism.
Is it sybilproof?
5 / 22
7. Negative Result
Assume reputation scores remain the same
under isomorphism.
Is it sybilproof?
Unfortunately, no. Attack strategy?
5 / 22
8. Negative Result
Assume reputation scores remain the same
under isomorphism.
Is it sybilproof?
Unfortunately, no. Attack strategy?
Answer: double the graph.
5 / 22
9. Positive Results (1/3)
General form of trust flow reputations:
r(x) = max trust(p)
Ptx
p∈Ptx
Notation:
t is pre-trusted node
Pxy is a family of disjoint paths from t to x
6 / 22
10. Positive Results (2/3)
Assumptions:
1
Extending path nonincreases the trust(p)
2
and trust are monotone to number of
paths and edges values, respectively
3
Splitting a path into two does not increase
value
7 / 22
11. Positive Results (2/3)
Assumptions:
1
Extending path nonincreases the trust(p)
2
and trust are monotone to number of
paths and edges values, respectively
3
Splitting a path into two does not increase
value
4
= max
7 / 22
13. Positive Results (3/3)
Under assumptions (1-3) sybil attack does not
increase adversary’s reputation
Under assumptions (1-4) sybil attack does not
increase adversary’s rank
8 / 22
14. Positive Results (3/3)
Under assumptions (1-3) sybil attack does not
increase adversary’s reputation
Under assumptions (1-4) sybil attack does not
increase adversary’s rank
Proof?
8 / 22
15. SybilGuard (1/2)
Assume number of attack edges is
A = o( n/ log n)
System is distributed, honest nodes follow
the same protocol
Can an honest node t identify (w.h.p.)
2A + 1 nodes in such a way that at most A
of them are powered by adversary?
9 / 22
16. SybilGuard (2/2)
For every node fix a bijective mapping
from in-edges to out-edges
Take a walk from t of length at most
n log n using bijection routing
At some point make a random switch,
than continue another n log n steps using
backwalk routing
Report a point. Repeat, until 2A + 1 points
are collected
10 / 22
17. SybilGuard (2/2)
For every node fix a bijective mapping
from in-edges to out-edges
Take a walk from t of length at most
n log n using bijection routing
At some point make a random switch,
than continue another n log n steps using
backwalk routing
Report a point. Repeat, until 2A + 1 points
are collected
Claim
w.h.p. at most A reported nodes are malicious
10 / 22
21. BlogRank
Any ideas how to rank blogs?
Why not just PageRank?
13 / 22
22. BlogRank
Any ideas how to rank blogs?
Why not just PageRank?
Wait a minute, for which graph?
13 / 22
23. BlogRank
Any ideas how to rank blogs?
Why not just PageRank?
Wait a minute, for which graph? Linked blogs:
Hyperlinks, blogrolls
Common commentors/authors, tags, co-references
to news
13 / 22
25. B2Rank
B2Rank(x) = BlogReputation × PostQuality
BlogReputation is computed in PageRank style
for blogroll graph with one change:
Blogroll links are weighted by activity level
(frequency of blogging and commenting)
14 / 22
26. B2Rank
B2Rank(x) = BlogReputation × PostQuality
BlogReputation is computed in PageRank style
for blogroll graph with one change:
Blogroll links are weighted by activity level
(frequency of blogging and commenting)
PostQuality is average for PageRank-style
score of blog posts
Post-to-post links are weighted by
referring post activity and time difference
14 / 22
28. EigenRumor (2/2)
Notation:
¯: reputation score for posts
r
¯ ¯
a, h: authority and hub scores for bloggers
P, E: provision and evaluation matrices
16 / 22
29. EigenRumor (2/2)
Notation:
¯: reputation score for posts
r
¯ ¯
a, h: authority and hub scores for bloggers
P, E: provision and evaluation matrices
¯
¯ = αPT a + (1 − α)ET h
r ¯
¯
a = P¯, h = E¯
¯ r r
16 / 22
30. EigenRumor (2/2)
Notation:
¯: reputation score for posts
r
¯ ¯
a, h: authority and hub scores for bloggers
P, E: provision and evaluation matrices
¯
¯ = αPT a + (1 − α)ET h
r ¯
¯
a = P¯, h = E¯
¯ r r
Solution: iterative algorithm for ¯:
r
T T
¯ = (αP P + (1 − α)E E)¯
r r
16 / 22
41. Challenges
Measurable objectives?
Model for input data?
Dynamic aspects of reputations?
Digg-style ranking?
Price of attack?
Ranking in social networks?
Ranking in RDF data?
Billion dollar question: how to avoid arms
race?
20 / 22
42. References
K. Fujimura, T. Inoue, M. Sugisaki
The EigenRumor Algorithm for Ranking Blogs
A. Kritikopoulos, M. Sideri, I. Varlamis
BlogRank: ranking weblogs based on connectivity and similarity features
M.A. Tayebi, S.M. Hashemi, A. Mohades
B2Rank: An Algorithm for Ranking Blogs Based on Behavioral Features
A. Cheng, E. Friedman
Sybilproof reputation mechanisms
H. Yu, M. Kaminsky, P.B. Gibbons, A, Flaxman
SybilGuard: defending against sybil attacks via social networks
P.A. Chirita, J. Diederich, W. Nejdl
MailRank: using ranking for spam detection
Z. Gyongyi, H. Garcia-Molina, J. Pedersen
Combating web spam with TrustRank
M. Dalal
Spam and popularity ratings for combating link spam
21 / 22