Neural Network-Based Reputation Model in a Distributed System * Weihua Song Computer Science Department Louisiana Tech University Ruston, LA 71270 wso003@latech.edu Vir V. Phoha Computer Science Department Louisiana Tech University Ruston, LA 71270 phoha@latech.edu Abstract Current centralized trust models are inappropriate to ap- ply in a large distributed multi-agent system, due to various evaluation models and partial observations in local level reputation management. This paper proposes a distributed reputation management structure, and develops a global reputation model. The global reputation model is a novel application of neural network techniques in distributed rep- utation evaluations. The experimental results showed that the model has robust performance under various estimation accuracy requirements. More important, the model is adap- tive to changes in distributed system structures and in local reputation evaluations. 1. Introduction Current reputation evaluations can be classified into two categories, central and personalized. Central reputation models make evaluations based on complete observations of online users’ trust behaviors. Personalized trust mod- els aggregate partial observations based on certain selection criteria and aggregation algorithms [2, 3]. However, current trust models do not provide a mecha- nism in managing users’ global reputations in a distributed system. This paper proposes a distributed master-slave rep- utation management structure. The distributed structure has advantages over a large and sparse central system in optimal local reputation management and in load balance of compu- tation time and memory storage. The paper also develops a global reputation model to derive global reputations from distributed local reputations. The model is adaptive (1) to heterogeneous local reputation models, and (2) to changes of a distributed system structure, users’ trust behaviors, and local reputation models etc. * This work is supported in part by the Army Research Office under Grant No. DAAD 19-01-1-0646 and by Louisiana Board of Regents under Grant LEQSF(2003-05)-RD-A-17. The rest of the paper is organized as follows. Section 2 proposes a distributed reputation management structure. Section 3 designs a neural network-based global reputation model. Section 4 presents experimental results. Section 5 concludes the paper. 2. Distributed Reputation Management Adoption of distributed reputation management in- creases reliability in certain aspects: (1) Reputations eval- uated by users of similar trust standards and opinions are more reliable and helpful in decision making. (2) There is no single reputation model optimal for various natures of online transactions. Distributed local communities have the flexibility to decide which reputation model best suits its community members’ interests (context dependent) and therefore have high impact on users’ online transaction de- cisions. (3) Distributed pairwise trust management is more efficient than centralized pairwise management in terms of memory space and data retrieval time. A central sys- tem of N users takes O(N 2 ) memory space to keep rat- ing records per reputation context if data structure arrays are used. However, online traders are usually very sparsely connected. Users of brand, price or quality preferences may never trade with quite a large portion of other users. If the same system can be divided into n (n can be very large in a sparse system) highly connected sub systems (assumed of equal size), it takes only O( N 2 n ) memory space. If linked lists are used instead, a central reputation system takes on average N 2 retrieval time for a rating, while the retrieval time is only N 2n on average in a distributed system. In addition, a distributed reputation system has privileges in balancing loads between global reputation management and distributed local reputation management. Figure 1 pro- poses a master-slave trust management structure. There are one global master agent and many distributed local slave agents. The master agent groups users (context dependent) to distributed local agents. If under the same trust context, a Proceedings of the IEEE International Conference on E-Commerce Technology 0-7695-2098-7/04 $20.00 © 2004 IEEE