reputation systems
TRANSCRIPT
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Gaming in Reputation systems
Anjan Goswami
June 9, 2011
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
What is Reputation System?
Computes a reputation score of entities in a domain.
Critical for many applications such as e-commerce.
Domain can often be treated as a graph or network.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Examples
Epinion (product reviews.)
Tripadvisor (hotel reviews.)
Google (search results.)
eBay (seller’s trust.)
Amazon (merchant’s trust.)
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Relationship with Recommandation systems
Conceptually similar.
Goals are different.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Objective
A brief introduction to reputation systems.
A discussion on page ranking.
A discussion on feedback ranking.
Attacks on reputation systems.
Making robust reputation systems.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
A naive page rank
πv =!
(w ,v)∈E
πwdw
!
v
πv = 1
πv ≥ 0 ∀v
dw is number of out links.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
A better page rank
πv = (1− ϵ)(!
(w ,v)∈E
πwdw
) +ϵ
N
!
v
πv = 1
πv ≥ 0 ∀v
N is total number of web pages.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Collusion in page rank
Number of hotels linking each other.
Page rank in the collusive community will be high.
Zero sum game between search engine companies and seocompanies.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Feedback based reputation system
The mechanism asks every buyer to provide feedback.
Every feedback is used to update the seller’s reputation.
A feedback score is computed based on some features relatedto feedbacks.
A Bayesian scheme can be used.
Example: eBay, RentACoder, Slashdot
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Features for reputation score
The percentage of positive feedback received by the user.
The number of unique users who left positive feedback.
The number of unique users who left negative feedback.
The total number of positive feedback received for alltransactions.
The average of the ratings for the additional dimensions,
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Bayesian reputation
The posterior probability of the reputation of a product oftype θ
Given the feedback qj ∈ Q
p(θ|qj) =p(qj |θ)p(θ)
p(qj)
p(qj) is the probability that buyer observes a signal qj ∈ Q
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Threats to Reputation systems
Sybil attacks (Bad mouthing or Ballot stuffing).
Bootstrap issues.
Presentation bias.
Attacks on underlying network. (DoS)
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Sybil Attack
Bad mouthing: Falsely produce negative feedback on others.
Ballot stuffing: Produce positive feedback on self.
Collusions.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Bootstrap issues
Disadvantage of new comers.
Managing reputation over time.
Cold start in recommendation systems.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Presentation bias
Bias on current reputation.
Conceptually same as positional bias in Search.
Lots of literature on positional bias models.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
P2p network
Symmetric: one in which, an elements reputation dependssolely on the topology of the trust graph, and not on thenaming or identity of nodes
Asymmetric: one in which, there are specifically trusted nodesfrom which all reputation values propagate
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Symmetric network
Symmetric: Sybil elements to create a copy of the existinggraph representing trust relationships.
cannot distinguish original nodes from the copies.
Not Sybil proof.
Example: Google page rank.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Asymmetric network
Asymmetric: each entity separately computes a trust valuealong their unique paths to every other identity in the system.
Trusted nodes cannot be impersonated, so no gain byduplicating the graph.
Sybil proof.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Sybil attack on search engine
Community of web sites having out links pointing to eachothers.
Many websites with same or nearly content but different url.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Sybil attack in online marketplaces
Bad mouthing or ballot stuffing.
Fake account and purchasing low priced items from a seller.
Writing positive review of the seller and negative review ofcompetitor.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
How to defend against Sybil attacks
Near duplicate detection in search.
Linked spam community detection.
Demoting reputation of linked spam sites.
Detect spam blogs or spam sites.
Find multiple low price purchase for a seller.
Bad buyer experience.
Introduction Topics Page rank Feedback based reputation Threats and Robust Reputation systems
Incentive compatible reputation
Mechanism design problem.
Design a system where gaming is costly.
Lots of literature on incentive compatible reputation.
Basic idea is to formulate an optimization problem.