| dc.description.abstract | In the field of multiagent systems, trust and reputation systems are intended to assist agents in finding trustworthy partners with whom to interact. Earlier work of ours identified in theory a number of security vulnerabilities in trust and reputation systems, weaknesses that might be exploited by malicious agents to bypass the protections offered by such systems. In this work, we begin by developing the TREET testbed, a simulation platform that allows for extensive evaluation and flexible experimentation with trust and reputation technologies. We use this testbed to experimentally validate the practicality and gravity of attacks against vulnerabilities. Of particular interest are attacks that are collusive in nature: groups of agents (coalitions) working together to improve their expected rewards. But the issue of coalitions is not unique to trust and reputation; rather, it cuts across a range of fields in multiagent systems and beyond. In some scenarios, coalitions may be unwanted or forbidden; in others they may be benign or even desirable. In this document, we propose a method for detecting coalitions and identifying coalition members, a capability that is likely to be valuable in many of the diverse fields where coalitions may be of interest. Our method makes use of clustering in benefit space (a high-dimensional space reflecting how agents benefit others in the system) in order to identify groups of agents who benefit similar sets of agents. A statistical technique is then used to identify which clusters contain coalitions. Experimentation using the TREET platform verifies the effectiveness of this approach. A series of enhancements to our method are also introduced, which improve the accuracy and robustness of the algorithm. To demonstrate how this broadly-applicable tool can be used to address domain-specific problems, we focus again on trust and reputation systems. We show how, by incorporating our work into one such system (the existing Beta Reputation System), we can provide resistance to collusion. We conclude with a detailed discussion of the value of our work for a wide range of environments, including a variety of multiagent systems and real-world settings. | en |
