How Social Structure Improves Distributed Reputation Systems - Three Hypotheses

Reputation systems provide an incentive for cooperation in artificial societies by keeping track of the behavior of autonomous entities. The self-organization of P2P systems demands for the distribution of the reputation system to the autonomous entities themselves. They may cooperate by issuing recommendations of other entities’ trustworthiness. The recipient of a recommendation has to assess its truthfulness and consistency before taking it into account. The current assessment methods are based on plausibility considerations that have several inherent limitations. In our previous work, we have suggested the application of non-repudiable tokens that overcome most of the limitations. However, there remain limitations that are not overcome or only partly overcome. Therefore, in this paper, we propose social structure as a complementary means of overcoming the remaining limitations of plausibility considerations. For this purpose, we examine the properties of social structure and discuss how distributed reputation systems can make use of them. This leads us to the formulation of three hypotheses of how social structure overcomes the limitations of plausibility considerations. In addition, it is discussed how the hypotheses can be tested. 1 The work done for this paper is partly funded by the German Research Community (DFG) in the context of the priority programs (SPP) no. 1140 and no. 1083. Simulative Test of the Three Hypotheses In this section, we present a simulative test of the three hypotheses that have been proposed in the previous section. First, we discuss the simulation environment that we used for these tests. Subsequently, we present and interpret the simulation results that were obtained for the respective hypotheses. Simulation Environment We have conducted the simulation in the DIANEmu simulator [1]. For the definition of the simulation environment, the benchmark group IBR2 has been applied [2]. The instances of the benchmark group differ in their definition of the network’s volatility and the ratio between cooperative and vicious agents. Cooperative agents never defect but they only enter into transactions with rather trustworthy agents. Vicious agents participate in every transactions and always defect. There are two types of vicious agents, i.e., uncooperative and colluding agents. In contrast to the uncooperative agents, colluders aim at manipulating the distributed reputation system by mutual praising. According to the analysis of Section ??, the Buddy System exploits best the potential of social structure. Therefore, we have chosen it as the representative of distributed reputation systems based on social structure. Simulation Results and Interpretation Hypothesis 1 (orientation for newcomers). For the corroboration of the first hypothesis, we have to show that the presence of social structure improves the performance of newcomers. At the early stages of the system, the agents have not yet established any relationships so that early newcomers cannot rely on the social structure. In contrast, late newcomers are provided an orientation by a well-established social structure. Therefore, we are able to measure the improvement of performance by comparing late newcomers with early newcomers. The simulation results are illustrated in Figure 1(a) for early newcomers (without social structure) and late newcomers (with social structure). It appears that the presence of social structure increases the performance of the newcomers by about 20%. Even in settings with large collusions (IBR2b: 30% of the population), with few cooperative agents (IBR2c: 20% of the population) or with few transactions (IBR2g: half as much as for IBR2a), social structure provides an orientation for newcomers. Hypothesis 2 (protection against collusions). For the corroboration of the second hypothesis, we have to demonstrate that the impact of mutual praising is marginalized by the presence of social structure. For this purpose, we distinguish between two types of colluders. The colluders may mutually establish buddy-relationships or they may resort to conventional recommendations in order to conduct mutual praising. The performance of these two types of colluders