UCDavis, ecs251 Fall 2007 11/29/2007Trust and Reputation System1 S. Felix Wu University of...

11/29/2007 Trust and Reputation System 1

UCDavis, ecs251Fall 2007

Trust and Reputation System

S. Felix WuUniversity of California, Davis

[email protected]://www.cs.ucdavis.edu/~wu/


UCDavis, ecs251Fall 2007 Computational Trust

representing a trust relationship between two directly communicating entities

Trust Attribute



• Trust Values– I “trust” him “50/50”.– I trust him “0.715”

• Partial Ordering Relationship– “I trust Alice more (than Bob)”– “I trust Alice more than the set threshold of

my spam mail filter”


UCDavis, ecs251Fall 2007 Trust Ordering

• Trust Ordering– I trust you, otherwise, I don’t.

• Information-based Ordering– I trust you, I don’t, or I don’t know based on

the information I have currently.– Dynamics and Uncertainty


UCDavis, ecs251Fall 2007 Policy & Delegation

• Policy:– If X trusts Y by Z, then A will trust B by C.– E.g.

• If Bank American will lend you $1M, then Washington Mutual will lend you $2M.


UCDavis, ecs251Fall 2007 Policy & Delegation

• Policy:– If X trusts Y by Z, then A will trust B by C.– E.g.

• If Bank American will lend you $1M, then Washington Mutual will lend you $2M.

– Trust means “Action and Risk”– Computational Trust needs to quantify the

actions and their associated risks.– It might be “Mutual Recursive” though…



• Direct DSL Link– Observing our direct neighbor’s behavior

• Indirect Sources in Social Network– Trust delegation– About a peer, may or may not be your direct

neighbor


UCDavis, ecs251Fall 2007 Trust in P2P

• The Service Provider provides a management system for trust and reputation– Google’s “PageRank”– Antivirus system– eBay’s seller reputation system– PKI

• P2P -- everything hopefully to be P2P– Decentralized model for trust


UCDavis, ecs251Fall 2007 Cheating & Incentives

• Selfish users in Gnutella and Bittorrent• eBay flaw seller ranking• Google page rank

• Selfishness or Reputation boost


UCDavis, ecs251Fall 2007 P2P Trust Model

• Less vulnerable?• Harder to implement? In a decentralized

setting?



• Problem: – Reduce inauthentic

files distributed by malicious peers on a P2P network.

• Motivation:

Problem

“Major record labels have launched an aggressive new guerrilla assault on the underground music networks, flooding online swapping services with bogus copies of popular songs.”

-Silicon Valley Weekly


UCDavis, ecs251Fall 2007 Problem

• Goal: To identify sources of inauthentic files and bias peers against downloading from them.

• Method: Give each peer a trust value based on its previous behavior.

0.9

0.1


UCDavis, ecs251Fall 2007 Some approaches

• Past History• Friends of Friends• EigenTrust• PeerTrust• TrustDavis


UCDavis, ecs251Fall 2007 Terminology

• Local trust value: cij. The opinion that peer i has of peer j, based on past experience.

• Global trust value: ti. The trust that the entire system places in peer i.

Peer 1

Peer 3

Peer 2

Peer 4

t4=0

t1=.3

t3=.5

t2=.2

C21=0.6

C23=0.7

C14=0.01

C12=0.3


UCDavis, ecs251Fall 2007 Local Trust Values

• Each time peer i downloads an authentic file from peer j, cij increases.

• Each time peer i downloads an inauthentic file from peer j, cij decreases.

Peer i Peer j

Cij=


UCDavis, ecs251Fall 2007 Normalizing Local Trust Values

• All cij non-negative

• ci1 + ci2 + . . . + cin = 1

Peer 2

Peer 1

Peer 4

C14=0.1

C12=0.9

Peer 2 Peer 4

Peer 1


UCDavis, ecs251Fall 2007 Local Trust Vector

• Local trust vector ci: contains all local trust values cij that peer i has of other peers j.

⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜

⎝

⎛

1.0

0

9.0

0

⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜

⎝

⎛

−−0

0

Peer 2

Peer 4

=

Peer 1

c1Peer 2

Peer 1

Peer 4

C14=0.1

C12=0.9


UCDavis, ecs251Fall 2007 Past history

• Each peer biases its choice of downloads using its own opinion vector ci.

• If it has had good past experience with peer j, it will be more likely to download from that peer.

• Problem: Each peer has limited past experience. Knows few other peers.

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

Peer 4

Peer 6

Peer 1

???

?

??


UCDavis, ecs251Fall 2007 Friends of Friends

• Ask for the opinions of the people who you trust.

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

Peer 4

Peer 6

Peer 1

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

Peer 2

Peer 8


UCDavis, ecs251Fall 2007 Friends of Friends

• Weight their opinions by your trust in them.

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

Peer 4

Peer 1

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

Peer 2

Peer 8

Peer 4


UCDavis, ecs251Fall 2007 The Math

∑ ⋅=j

jkij cccik

'

Ask your friends j

What they think of peer k.

And weight each friend’s opinion by how

much you trust him.

TC'ic ic=

.1

.5 0 0 0.2

0 .2 0 .3 0 .5 .1 0 0 0

.1

.3

.2

.3

.1

.1

.2 =


UCDavis, ecs251Fall 2007 Problem with Friends

• Either you know a lot of friends, in which case, you have to compute and store many values.

• Or, you have few friends, in which case you won’t know many peers, even after asking your friends.

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0


UCDavis, ecs251Fall 2007 Dual Goal

• We want each peer to:– Know all peers.– Perform minimal computation (and storage).


UCDavis, ecs251Fall 2007 Knowing All Peers

• Ask your friends: t=CTci.

• Ask their friends: t=(CT)2ci.

• Keep asking until the cows come home: t=(CT)nci.

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛−−

0

0

0

0

0

0

⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟

⎠

⎞

⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜

⎝

⎛

−−

0

0

0

0

0

0


UCDavis, ecs251Fall 2007 Minimal Computation

• Luckily, the trust vector t, if computed in this manner, converges to the same thing for every peer!

• Therefore, each peer doesn’t have to store and compute its own trust vector. The whole network can cooperate to store and compute t.


UCDavis, ecs251Fall 2007 Non-distributed Algorithm

• Initialize:

• Repeat until convergence:

(k)T1)(k tCt =+

T(0)t ⎥⎦

⎤⎢⎣

⎡=nn1

...1


UCDavis, ecs251Fall 2007 Distributed Algorithm

• No central authority to store and compute t.

• Each peer i holds its own opinions ci.

• For now, let’s ignore questions of lying, and let each peer store and compute its own trust value.

)()(11

)1( ... knni

ki

ki tctct ++=+

.1

.5 0 0 0.2

0 .2 0 .3 0 .5 .1 0 0 0

.1

.3

.2

.3

.1

.1

.2 =


UCDavis, ecs251Fall 2007 Distributed Algorithm

For each peer i { -First, ask peers who know you for their opinions of you. -Repeat until convergence {

-Compute current trust value: ti(k+1) = c1j t1

(k) +…+ cnj tn(k)

-Send your opinion cij and trust value ti(k) to your

acquaintances.-Wait for the peers who know you to send you their trust values and opinions.

}}


UCDavis, ecs251Fall 2007 Probabilistic Interpretation


UCDavis, ecs251Fall 2007 Malicious Collectives


UCDavis, ecs251Fall 2007 Pre-trusted Peers

• Battling Malicious Collectives

• Inactive Peers• Incorporating

heuristic notions of trust

• Convergence Rate


UCDavis, ecs251Fall 2007 Secure Score Management

• Two basic ideas:– Instead of having a

peer compute and store its own score, have another peer compute and store its score.

– Have multiple score managers who vote on a peer’s score.

M

M

M

M

Score Manager

Score Managers

?

?

??

Distributed Hash Table


UCDavis, ecs251Fall 2007 PeerTrust System Architecture

P1

P3

P4

P2P Network

Trust Data

Data Locator

Feedback Submission

Trust Evaluation

Trust Manager

P5

P6

P2P Network

P2


UCDavis, ecs251Fall 2007 How to use the trust values ti

• When you get responses from multiple peers:– Deterministic: Choose the one with highest

trust value.– Probabilistic: Choose a peer with probability

proportional to its trust value.


UCDavis, ecs251Fall 2007 Load Distribution

Deterministic Download Choice

Probabilistic Download Choice


UCDavis, ecs251Fall 2007 Threat Scenarios

• Malicious Individuals– Always provide

inauthentic files.

• Malicious Collective– Always provide

inauthentic files.– Know each other. Give

each other good opinions, and give other peers bad opinions.


UCDavis, ecs251Fall 2007 More Threat Scenarios

• Camouflaged Collective– Provide authentic files

some of the time to trick good peers into giving them good opinions.

• Malicious Spies– Some members of the

collective give good files all the time, but give good opinions to malicious peers.


UCDavis, ecs251Fall 2007 Malicious Individuals


UCDavis, ecs251Fall 2007 Malicious Collective


UCDavis, ecs251Fall 2007 Camouflaged Collective


UCDavis, ecs251Fall 2007 P2P Electronic Communities


UCDavis, ecs251Fall 2007 Motivation



• Should we buy?• How do we decide?



• Should we buy?• How do we decide?

• What we want:– accurately estimate risk of default– minimize the risk of default– minimize losses due to pseudonym change– avoid trusting a centralized authority

• How do we achieve these goals?



• TrustDavis is a reputation system that realizes these goals.

• It recasts these goals as the following properties:



1. Agents can accurately estimate risk– Third parties provide accurate ratings

2. Honest buyer/seller avoids risk (if possible)– Insure transactions

3. No advantage in obtaining multiple identities – Agents can cope with pseudonym change

4. No need to trust a centralized authority– No centralized services needed



Incentive Compatibility:

Each player should have incentives to perform the actions that enable the system to achieve a desired global outcome.



1. Agents can accurately estimate risk– Third parties provide accurate ratings

2. Honest buyer/seller avoids risk (if possible)– Insure transactions

3. No advantage in obtaining multiple identities – Agents can cope with pseudonym change

4. No need to trust a centralized authority– No centralized services needed

Incentive Compatibility!



A Reference is:Acceptance of Limited Liability.

$100

BAC



1. Agents can accurately estimate risk– Third parties provide accurate ratings– Parties are liable for the references they provide

2. Honest buyer/seller avoids risk (if possible)– Insure transactions– Buyers/sellers pay for references to insure their

transactions3. No advantage in obtaining multiple identities

– Agents can cope with pseudonym change– References are issued only to trusted identities

4. No need to trust a centralized authority– No centralized services needed– Anyone can issue a reference

Use References!


UCDavis, ecs251Fall 2007 Outline

• TrustDavis leverages social networks

• For now, examples assume No False Claims (NFC)

• The use of TrustDavis does NOT preclude trade outside the system.


UCDavis, ecs251Fall 2007 Paying for References

150

150100

50

50



• vb wants to buy three shirts.• Shirts cost $100 each from a

trustworthy seller• Unknown seller offers shirts

for $50 each (but maybe they are only worth $25).

• vb would risk 3 x $50 = $150 in the transaction

• vb can borrow and lend money at rate r=1.25 through the period of the transaction

For $30, vb can insure herself!

Paying for References

How much is vb willing to pay to insure the transaction? (No riskless profitable arbitrage criterion)

Example:

$100 each

Trust-me.com

Blowout SALE!

$50 each!$150!



To insure herself vb buys the shirts and a hedging portfolio as follows:

1. Instead of buying 3 shirts for $50 each she buys only 2, saving $50.

2. The buyer, vb , adds $30 of her own money and lends the resulting $80 at rate r = 1.25.



On Success:

– vb obtains $100 from the loan and buysthe 3rd shirt

On failure:

– vb sells the two shirts for $25 each

– gets $100 from the loan. – She obtains a total of $150

Thus, vb can insure herself for $30.


UCDavis, ecs251Fall 2007 Selling References



Seen as an investment…

On Success the ROI is:

On failure the ROI is:

If repeated many times the insurer may go bankrupt. Assume the insurer has W dollars available to insure this transaction.

K

C

K

CK+=

+1

K

C



Insurer maximizes the expected value of the growth rate of capital (Kelly Criterion).

For given:– probability of failure p,– a desired growth rate of capital R; and,– fraction of the total funds W being risked in a transaction.

The insurer can obtain a lower bound on the premium C.

⎪⎭

⎪⎬

⎫

⎪⎩

⎪⎨

⎧

⎥⎦

⎤⎢⎣

⎡=

nn

W

WER

1

0

log



Insured Value as a fraction of total funds – f

Co

st/In

sure

d V

alu

e –

C/K

Minimum Return/Risk Ration for Different Failure Probabilities


UCDavis, ecs251Fall 2007 A Non-Exploitable Strategy

Two Scenarios:• No False Claims - NFC• With False Claims - FC

False claims only change the probability p.We can incorporate the cost of verification.

Key Idea:

Save part of the money obtained in successful transactions in excess of the opportunity cost.



Example.

The buyer, vb, has $190 to spend on 1 of 3 options:

1. Buying 3 shirts from an unknown seller for $50 each and insuring the transaction for $40. She values each shirt at $100.

2. Buying 2 pairs of shoes from a reliable retailer for $70 each. She thinks each pair is worth $90.

3. Buying 1 game console for $150, from a reliable online shop. She values the console at $240.



vb’s valuation for each of the 3 options is:

1. Shirts: 100 x 3 + 0 (no cash leftover) = $300

2. Pairs of Shoes: 90 x 2 + 50 (cash) = $230

3. Console: 240 x 1 + 40 (cash) = $280

Gains in excess of the opportunity cost are:300-280=$20.

Part of these $20 should be saved to insure future transactions.



The Strategy:

1. Initially only provide references to known agents or those that leave a security deposit.

2. Insure all trade through references provided by trusted agents.

3. Do not provide more insurance than you can recover. Charge at least the lower bound for providing a reference.

4. Save part of the money received “in excess of the opportunity cost”.



150

150100

50

50

50

OK!$10 saved to

provide future insurance 10

Failed!Payment made

automatically by v1



• Motivation• The Model

– Buying references– Selling references

• A Non-Exploitable Strategy• Future Work• Conclusion

– Key ideas


UCDavis, ecs251Fall 2007 Future Work

• Simulation– sensitivity to estimates of p– growth rate of capital– dynamic behavior

• Price Negotiation– should avoid “double spending” problem– fair distribution among insurers of the premium paid



• Motivation• The Model

– Buying references– Selling references

• A Non-Exploitable Strategy• Future Work• Conclusion

– Key ideas


UCDavis, ecs251Fall 2007 Conclusion

TrustDavis provides:• Accurate Ratings• Non-exploitable strategy for honest

agents• Pseudonym change tolerance• Decentralized infrastructure

Through the use of References.


UCDavis, ecs251Fall 2007 Conclusion

Key Ideas:

• Incentive Compatibility– Incentive to accurately rate– Incentive to insure– No incentive to change pseudonym

• Saving gains in excess of the opportunity cost to insure future transactions.


UCDavis, ecs251Fall 2007 The End

UCDavis, ecs251 Fall 2007 11/29/2007Trust and Reputation System1 S. Felix Wu University of...

Documents

Transcript of UCDavis, ecs251 Fall 2007 11/29/2007Trust and Reputation System1 S. Felix Wu University of...