Defending Against Sybil Attacks Defending Against Sybil Attacks via Social Networksvia Social Networks
Haifeng Yu
School of Computing
National University of Singapore
Haifeng Yu, National University of Singapore 2
AcknowledgmentsAcknowledgments
Talk based on three papers [SIGCOMM’06, ToN’08] (SybilGuard)
[IEEE S&P’08] (SybilLimit)
Available on my homepage – google my name
Co-authors: Phillip B. Gibbons
Michael Kaminsky
Feng Xiao
Abie Flaxman
Haifeng Yu, National University of Singapore 3
Background: Sybil AttackBackground: Sybil Attack
Sybil attack: Single user pretends many fake/sybil identities I.e., Creating multiple accounts
Already observed in real-world p2p systems
Sybil identities can become a large fraction of all identities
launchsybilattack
honest
malicious
Haifeng Yu, National University of Singapore 4
Background: Sybil AttackBackground: Sybil Attack
Enables malicious users to easily “out-vote” honest users Byzantine consensus – exceed the 1/3 threshold
Majority voting – cast more than one vote
DHT – control a large portion of the ring
Recommendation systems – manipulate the recommendations
Haifeng Yu, National University of Singapore 5
Background: Defending Against Sybil AttackBackground: Defending Against Sybil Attack Using trusted central authority to tie identities to
human beings – not always desirable
Much harder without a trusted central authority [Douceur’02] Resource challenges not sufficient
IP address-based approach not sufficient
Widely considered as real & challenging: Over 40 papers acknowledging the problem of sybil
attack, without having a distributed solution
Haifeng Yu, National University of Singapore 6
SybilGuard / SybilLimit Basic Insight: SybilGuard / SybilLimit Basic Insight: Leveraging Social NetworksLeveraging Social Networks
Nodes = identities
Undirected edges = strong mutual trust E.g., colleagues,
relatives in real-world
Not online friends!
SybilGuard / SybilLimit is the first to use social networks for thwarting sybil attacks with provable guarantees.
Haifeng Yu, National University of Singapore 7
SybilGuard / SybilLimit Basic InsightSybilGuard / SybilLimit Basic Insight
malicioususers
honestnodes
Observation: Adversary cannot create extra edges between honest nodes and sybil nodes
attack edges
n honest users: One identity/node each
Malicious users: Multiple identities each (sybil nodes)
sybil nodes
sybil nodes may collude – the adversary
Haifeng Yu, National University of Singapore 8
SybilGuard/SybilLimit Basic InsightSybilGuard/SybilLimit Basic Insight
honest nodes sybil nodes
Dis-proportionally small cut disconnecting a large number of identities
But cannot search brute-force…attack
edges
Haifeng Yu, National University of Singapore 9
SybilGuard / SybilLimit End GuaranteesSybilGuard / SybilLimit End Guarantees
Completely decentralized
Enables any given verifier node to decide whether to accept any given suspect node Accept: Provide service to / receive service from
Ideally: Accept and only accept honest nodes – unfortunately not possible
SybilGuard / SybilLimit provably Bound # of accepted sybil nodes (w.h.p.)
Accept all honest nodes except a small fraction (w.h.p.)
Haifeng Yu, National University of Singapore 10
Example Application ScenariosExample Application Scenarios
If # of sybil nodes accepted
Then applications can do
< n majority voting
< n/2 byzantine consensus
< n/c for some constant c secure DHT [Awerbuch’06, Castro’02,
Fiat’05]
… …
Haifeng Yu, National University of Singapore 11
total number of attack edges
SybilGuard [SIGCOMM’06]
SybilLimit [Oakland’08]
nnOg log/ )log( nn )(log n
)(log nunbounded
# sybil nodes accepted (smaller is better) per attack edge
nn log/ nnO log/
g between
and
g
~2000 ~10
~10
SybilGuard vs. SybilLimitSybilGuard vs. SybilLimit
We also prove that SybilLimit is away from optimal)(log nO
Haifeng Yu, National University of Singapore 12
OutlineOutline
Motivation, basic insight, and end guarantees
SybilLimit design Will focus on intuition
Evaluation results on real-world social networks
Haifeng Yu, National University of Singapore 13
Cryptographic KeysCryptographic Keys
Each edge in social network corresponds to a symmetric edge key Established out of band
Each node (honest or sybil) has a locally generated public/private key pair “Identity”: V accepts S = V accepts S’s public key KS
When running SybilLimit, every suspect S is allowed to “register” KS on some other nodes
Haifeng Yu, National University of Singapore 14
SybilLimit: Strawman Design – Step 1SybilLimit: Strawman Design – Step 1
Ensure that sybil nodes (collectively) register only on limited number of honest nodes
Still provide enough “registration opportunities” for honest nodes
sybil regionhonest region
K: registered keys of sybil nodes
K K
K
KK
K
K K
K
K
K
K
K
KK K
K: registered keys of honest nodes
Haifeng Yu, National University of Singapore 15
SybilLimit: Strawman Design – Step 2SybilLimit: Strawman Design – Step 2
Accept S iff KS is
register on sufficiently many honest nodes
Without knowing where the honest region is !
Circular design? We can break this circle…
K K
K
KK
K
K K
K
K
K
K
K
KK K
sybil regionhonest region
K: registered keys of sybil nodes
K: registered keys of honest nodes
Haifeng Yu, National University of Singapore 16
Three Interrelated Key TechniquesThree Interrelated Key Techniques
Technique 1: Use the tails of random routes for registration Will achieve Step 1
SybilGuard novelty: Random routes
SybilLimit novelty: The use of tails
SybilLimit novelty: The use of multiple independent instances of shorter random routes
Haifeng Yu, National University of Singapore 17
Three Interrelated Key TechniquesThree Interrelated Key Techniques Technique 2: Use intersection condition and
balance condition to verify suspects Will break the circular design and achieve Step 2
SybilGuard novelty: Intersection on nodes
SybilLimit novelty: Intersection on edges
SybilLimit novelty: Balance condition
Technique 3: Use benchmarking technique to estimate unknown parameters Breaks another seemingly circular design…
SybilLimit novelty: Benchmarking technique
Haifeng Yu, National University of Singapore 18
Random 1 to 1 mapping between incoming edge and outgoing edge
Random Route: ConvergenceRandom Route: Convergence
a db ac bd c
d ee df f
a
b
c
d e
f
randomized
routing table
Using routing table gives Convergence Property:
Routes merge if crossing the same edge
Haifeng Yu, National University of Singapore 19
Securely Registering Public KeysSecurely Registering Public Keys
All random routes in SybilLimit are of length w All nodes know w
Nodes communicate via authenticated channels
A B C D
To register KA, A initiates a random route (assuming w = 3)
i = 1
KA
i = 2
KA
i = 3
KA
i = 3
KA
record KA
under name “CD”
edge “CD” is the tail of A’s random route
Haifeng Yu, National University of Singapore 20
Tails of Sybil SuspectsTails of Sybil Suspects Imagine that every sybil suspect initiates a
random route from itself
total 1 tainted tail
honestnodes
sybilnodes
tainted tail
Haifeng Yu, National University of Singapore 21
Counting The Number of Tainted TailsCounting The Number of Tainted Tails
Claim: There are at most w tainted tails per attack edge Proof: By the Convergence property
Regardless of whether sybil nodes follow the protocol
honestnodes
sybilnodes
attack edge
Haifeng Yu, National University of Singapore 22
Back to the Strawman Design Step 1Back to the Strawman Design Step 1
# of K ’s gw Independent of # sybil
nodes
# of K ’s n – gw From “backtrace-ability”
property of random routes
See paper…
honest region
K
K
K
K
K
K
KStep 1 achieved !
K: registered keys of sybil nodes
K: registered keys of honest nodes
Haifeng Yu, National University of Singapore 23
Independent InstancesIndependent Instances
SybilLimit uses independent instances of the registration protocol m: # of edges in the honest region
Number of K’s:
Number of K’s:
Goal: Accept S iff KS is registered on
tails in the honest region Sybil suspects accepted:
Honest suspects accepted:
m
mwgn )(
mwg
m
wgn
wg
Haifeng Yu, National University of Singapore 24
Three TechniquesThree Techniques Technique 1: Use novel random routes to
register public keys Will achieve Step 1
Technique 2: Use intersection condition and balance condition to verify suspects Challenge: SybilLimit does not know which region is
the honest region
Technique 3: Use benchmarking technique to estimate unknown parameters
Haifeng Yu, National University of Singapore 25
The Intersection ConditionThe Intersection Condition
Verifier V obtains tails by doing random routes of length w Using different instances – see paper…
Some tails are in the sybil region – ignore for now…
S satisfies intersection condition if: S’s and V’s tails intersect
S’s public key is registered with the intersecting tail
m m
Haifeng Yu, National University of Singapore 26
4. Is KS registered?
Intersection Condition: Verification ProcedureIntersection Condition: Verification Procedure
VS
1. request S’s set of tails AB
CDEF
F
2. I have three tails
AB; CD; EF
3.common tail: EF
5. Yes.4 messages involved
S satisfies intersection condition
Haifeng Yu, National University of Singapore 27
Leveraging Known Random Walk TheoryLeveraging Known Random Walk Theory
(Approximate) Theorem: If w is roughly the mixing time of the social network,
then all tails (V’s and S’s) are roughly uniformly random edges
If social networks have mixing time, then
)(log nO)(log nOw
Haifeng Yu, National University of Singapore 28
Leveraging a Sharp DistributionLeveraging a Sharp Distribution
Assuming V has tails in the honest region
1.0
0
Intersection prob p
# of S’s tails in honest region
m
m
1p
m
0pBirthday paradox
m
Help to bound # of sybil nodes accepted
This is why SybilLimit does edge intersection …
Haifeng Yu, National University of Singapore 29
Back to the Strawman Design Step 2Back to the Strawman Design Step 2
Accept S iff KS is
register on sufficiently many honest nodes
“Sufficiently many” =
Intersection occurs iff S has tails in the honest region
K K
K
KK
K
K K
K
K
K
K
K
KK K
sybil regionhonest region
K: registered keys of sybil nodes
K: registered keys of honest nodes
m
m
Haifeng Yu, National University of Singapore 30
Omitted Challenges …Omitted Challenges …
Some of V’s tails are in the sybil region We do not know which tails are in the sybil region
Balance condition – hardest part to prove in SybilLimit…
Adversary has many strategies to allocated the tainted tails…
Tainted tails are not uniformly random…
See paper for details…
Haifeng Yu, National University of Singapore 31
Three Interrelated Key TechniquesThree Interrelated Key Techniques
Technique 1: Random routes
Technique 2: Intersection condition and balance condition
Technique 3: Novel and counter-intuitive benchmarking technique Avoids another seemingly circular design…
See paper…
Claims on near-optimality: See paper…
Haifeng Yu, National University of Singapore 32
Performance AspectsPerformance Aspects Random routes are performed only once
Re-do only when social network changes – infrequently
Can be done incrementally
Doing random routes is not time-critical Only delays a new suspect being accepted
Churn is a non-problem…
Verification involves O(1) messages
See paper…
Haifeng Yu, National University of Singapore 33
OutlineOutline
Motivation, basic insight, and end guarantees
SybilLimit design
Evaluation results on real-world social networks
Haifeng Yu, National University of Singapore 34
Validation on Real-World Social NetworksValidation on Real-World Social Networks
SybilGuard / SybilLimit assumption: Honest nodes are not behind disproportionally small cuts Rigorously: Social networks (without sybil nodes) have
small mixing time
Mixing time affects # sybil nodes accepted
Synthetic social networks – proof in [SIGCOMM’06]
Real-world social networks? Social communities, social groups, ….
Haifeng Yu, National University of Singapore 35
Simulation SetupSimulation Setup
We experiment with: Different number and placement of attack edges
Different graph sizes -- full size to 100-node sub-graphs
Sybil attackers use the optimal strategy
# nodes # edges
Friendster 0.9M 7.8M
Livejournal 0.9M 8.7M
DBLP 0.1M 0.6M
Crawled online social networks used in experiments
Haifeng Yu, National University of Singapore 36
Brief Summary of Simulation ResultsBrief Summary of Simulation Results
In all cases we experimented with:
Average honest verifier accepts ~95% of all honest suspects
Average honest suspect is accepted by ~95% of all honest verifiers
# sybil nodes accepted: ~10 per attack edge for Friendster and LiveJournal
~15 per attack edge for DBLP
Haifeng Yu, National University of Singapore 37
Other Social Networks?Other Social Networks?
Other social networks likely to have small mixing time too (DBLP as a worst-case)
What if the mixing time is large? Graceful degradation of SybilLimit’s guarantees --
Accept more sybil nodes
Haifeng Yu, National University of Singapore 38
ConclusionsConclusions
Sybil attack: Widely considered as a real and challenging problem
SybilLimit: Fully decentralized defense protocol based on social networks Provable near-optimal guarantees
Experimental validation on real-world social networks
Future work: Implement SybilLimit with real apps
Haifeng Yu, National University of Singapore 39
Post Doc OpeningPost Doc Opening NUS: Ranked 31st globally by Newsweek
E.g., we have 11 SIGMOD papers in 2008
I have post doc opening in distributed systems and distributed algorithms Minimum 1 year, renewable up to multiple years
2 years funding already committed
Main job duty: Publish in top venues Help you to build up track record for career after post doc
Salary: Comparable (if not better) than US post docs Singapore living cost and tax are lower than US
Contact me to inquire or apply – google my name