Attacks and defense techniques for reputation systems

[jghms]

We try to build a secure reputation system so it seems valuable to study surveys on possible attacks and defenses for reputation systems. First and foremost the Sybil attack is what I need to understand.

[jghms]

I focussed on this paper

[jghms]

Reputation systems have three important dimensions: Formulation (metric of reputation, aggregation of votes), Calculation (algorithm for aggregation and distribution) and Dissemination (spreading of results).

[jghms]

Attack types

• Self-promoting - falsely increasing one's reputation score, for example using the Sybil attack
• Whitewashing - restoring reputation after abusing the system to escape consequences
• Slandering - manipulate the reputation of other nodes on the system
• Orchestrated - combination of the above methods
• Denial of service - preventing of calculation or dissemination of reputation

[jghms]

Design choices for reputation formulation

• source of information
  • manual - nodes manually give feedback on interactions like a review on reputation networks
  • direct automatic - a record of a successful transaction can be an automatic source of information
  • indirect automatic - a record of an observed successful transaction between two other nodes
• information type
  • positive events only - bad behavior cannot be tagged
  • negative events only - good behavior cannot be tagged
  • hybrid (both positive and negative)
• Reputation metric
  • binary
  • discrete
  • continuous

[jghms]

Preventing sybil attacks

• increase cost for creating an identity
• flow based approaches

[jghms]

Preventing false rumors

• connect all ratings to real interactions
• dishonest feedback filter
• only direct interactions so you cannot spread false rumors (very limiting because indirect reputation helps with information asymmetry)
• use of pre-trusted nodes (EigenTrust)

[jghms]

Preventing short-term abuse

• newcomers start from 0 and are only trusted after gaining reputation through good behavior
• start with a negative score, "pay your dues" before being citizen of the system
• weight of historical trust fades, making short-term actions more important, dangerous if false rumors are possible

[jghms]

Reputation systems

• CORE, Michiardi and Molva, 2002
• EigenTrust, Kamvar et al., 2003
• TrustGuard, Srivatsa et al., 2005
• Scrivener, Nandi et al., 2005
• P2PRep, Aringhieri et al., 2006
• Credence, Walsh and Sirer, 2006

[jghms]

Scrivener

• relies on a neoclassical economics rational user model, economical incentive to act correctly
• no global reputation but local, peer-2-peer credit balance
• requesting a file gives the other agent a small positive credit while one-self has 0 credit
• if a node has 0 credit it cannot download without first uploading to the other agent
• mostly to defend against free-riders

[jghms]

P2PRep

• fuzzy reputation system that aggregates information from direct interactions and network opinion into a reputation and trust score