200 likes | 741 Views
REPUTATION-BASED TRUST MODELLING. Gayatri Swamynathan CS290F, 12/2/04. OUTLINE. Quick Overview Clarifications Project Changes The Model Performance Metrics Conclusions and Future Work. OVERVIEW. Trust Management: any mechanism that helps establish trust (or distrust) between peers
E N D
REPUTATION-BASED TRUST MODELLING Gayatri Swamynathan CS290F, 12/2/04
OUTLINE • Quick Overview • Clarifications • Project Changes • The Model • Performance Metrics • Conclusions and Future Work
OVERVIEW • Trust Management: any mechanism that helps establish trust (or distrust) between peers • Reputation is a measure that is derived from direct or indirect knowledge of earlier interactions of peers and is used to access the level of trust a peer puts into another • Reputation-based Trust Management : A Risk Management Technique
Trust: The Notion of Context • Trusting a peer to • Provide good service (here, files) • Provide good referrals/opinions • Malicious (false positives/negatives) • Incompatible viewpoints
Some Project Changes • Decentralized network (more generic) • Not just a survey • Implementing a trust model to understand the benefits of using reputation
OUTLINE • Quick Overview • Clarifications • Project Changes • The Model • Performance Metrics • Conclusions and Future Work
The Model: File Transfer Bootstrap: File Holders Random File Requests Generator 2 List of File X Providers 1 3 Peer requests file X File Transfer 5 Process Trust Values to choose the best peer Post-Transaction update 4 6 Local Trust Table
The Model: Representing Trust • Data Structures to represent Trust • ServiceTrust (st) • opinionTrust (op) • firstHand (fh) to represent direct-interaction observations • Tolerance Thresholds • serviceThreshold • values lower than this indicate untrustworthiness • If st(i,j) > serviceThreshold, interact! • If no serviceTrust value known, trust strangers! • opinionThreshold
The Model: Transfer of Trust Node i receives fh(k,j) where: firstHand information on Node j generated by Node k, post transaction Case1: If op(i,k) > opinionThreshold { Accept fh(k,j) Modify st(i,j) Add k to goodOps(j)/badOps(j) } Case2: If op(i,k) > opinionThreshold , but st(i,j)≈ fh(k,j) Do NotAccept fh(k,j) Modify op(i,k) }
The Model: Transfer of Trust Case3: If op(i,k) < opinionThreshold { Do Not Accept fh(k,j) } Case4: If op(i,k) < opinionThreshold , but st(i,j)≈ fh(k,j) Accept fh(k,j) Modify op(i,k) Add k to goodOps/badOps list } Case5: No Opinion Values Known: trust stranger’s opinions!
The Model But wait… Node i now interacts with Node j (i.e. st(i,j) > serviceThreshold) If the interaction is bad, Node i checks goodOps(j) and reduces opinionTrust values of all the nodes that gave a thumbs-up to Node j !!
Simulation Setting: Topology • Decentralized Network (10 nodes – 25 nodes) • Stanford GraphBase • Platform for general graph representation and manipulation • GT-ITM (Georgia Tech Internetwork Topology Model) • Creation and analysis of graph models of network topology • Implementation in NS2 and C++ • Peer Agent • FTP Agent
Simulation Setting: Sample Topology • Parameters: • Number of nodes • Probability of an edge from a node
OUTLINE • Quick Overview • Clarifications • Project Changes • The Model • Performance Metrics • Conclusions and Future Work
Performance Metrics • Mean Time to detect malicious behavior • with and without reputation-based trust model • different tolerance thresholds • Overheads: • Storage • ~20 bytes per trust-table entry for a 10-node network !! • Timestamps? • Control Messages • UDP packets
Performance Metrics: Detecting Malicious Behavior With and Without Reputation (threshold values = 0.5)
Performance Metrics: Detecting Malicious Behavior Different Tolerance thresholds (t=0.25, t=0.5)
Conclusions • Decentralized networks with Reputation-based trust mechanisms help systems work better. • Future Work • Post Transaction Analysis of Requesting-Peer • Collusion • Reputation-history similar to Credit History