On The Policy Description And Quantification Model For Trust Management

The internet based information system is becoming increasingly important for the human society. The e-business, e-government and e-society applications are widely deployed on the internet, and various new computing techniques such as cloud, grid and peer-to-peer computing, have made the internet-based applications more convient and efficient. Althougth the internet can result in great benefit, it also causes huge challenge for both the trust in the human society and the trust between human and machaine. With the unfamiliar client and unknown service quality, the trust risk for an internet application has been increased considerably. Hence, a trustworth internet environment becomes one of the preconditions for further improvement of internet based applications.According to the requirements of the internet based applications, we classify the trust management system into four trust levels: the ID-based trust, the utinity-oriented trust, the reliability-oriented trust and the robustness-oriented trust. For the trust in the first three levels, we put forward four closely related questions: how to design a powerful policy language for the ID-based trust management, how to get a comprehensive and robust reputation model for a service utility, how to combine the policy-based and reputation-based trust management, and how to apply the trust in a complicated network system to optimize its reliability oriented trust. To solve these four questions, our work in this paper can be briefly introducd as following:(1) Distributed Proving Oriented Language and Method for Trust NegotiationMost existing trust policy languages cannot simultaneously support the following important characteristics: distributed trust proving, complicated access control definition and negotiation-related constraints. This paper presents the RTP (role-based trust proving) policy language for distributed trust proving and negotiation. The main contributions of RTP include: through entending the RT language, it can define complicated roles; with the predicate lsign, it can define a logic role and support the distributed trust proving; with the predicate release, it can protect the policy's sensitive information; and it can avoid the unrelated credential fetching with the help of the negotiation heuristics. Both the syntax and semantics of RTP are introduced. In addition, we proved the soundness and completeness of RTP through the definatation of its inference rules. Based on RTP language, we design a distributed trust proving and negotiation algorithm, which can carry out an efficient trust construction through local trust negotiation or remote trust proving. We also demonstrate the soundness and completeness of DPN algorithm based on the trust graph. Our experiments show that the DPN algorithm outperforms the traditional trust negotiation in terms of both time and number of credential transfers. (2) A Comprehensive and Robust General Trust Model for Reputation EvaluationReputation-based trust management is becoming increasingly important in distributed systems. Although reputation is a prediction about the trust probability, most existing work cannot assess the reputation prediction variance. Moreover, the summation method is widely used for feedback aggregation, but it is vulnerable to malicious feedbacks and difficult to be protected. This paper presents a general trust model RLM, whose highlight is the use of reputation predication variance to give a more comprehensive and robust reputation evaluation. Concretely, we define the reputation by two attributes: reputation value and reputation prediction variance. For feedback aggregation, we introduce the novel Kalman aggregation method, which can support robust trust evaluation. To make the model robust, we design the Expectation Maximization algorithm to mitigate the influence of a malicious feedback, and further apply the hypothesis test method to resist the malicious feedback. Through theory analysis, we demonstrate the robustness of our design. Our experiments show that RLM model can effectively capture the reputation and its prediction variance. Compared with some other typical trust models, RLM can give a more accurate reputation prediction. Moreover, it has a high robustness under the attack of malicious feedbacks.(3) A Combined Trust Management with Policies and Reputation ValueMost existing role-based policy languages define trust as a Boolean-role relationship [0,1], which cannot support fine granularity access control. In addition, the current static role management system cannot track the authorized role usage to defend the malicious role behavior. In this paper, we propose a trust management system RTE which can combine the advantages of both the policy-based and reputation-based trust management. The policy language of RTE can support the trust value management through adding the parameter of reputation value. We compute the reputation value by the trust experience and trust recommadetion, so that RTE can give a fine granularity access control. In addition, through defining a trust aggregation operator for the social network, RTE can track the role behavior, hence, it can dynamically manage the role and defend the malicious role behavior. Both the synax and the inference rules of the policy language in RTE are introduced, and we give a demonstrating exaple of how to manage the trust with RTE.(4) Reliability-oriented Reputation and Its Application in Workflow OptimizationMost existing work used the reputation to make a decision for a simple application with only one transaction involved. For a workflow application, which is composed of many sequential or paremerall tasks, providing a reliable scheduling based on resource reliability evaluation is becoming increasingly important. Most existing reputation models used for reliability evaluation ignore the task runtime influence. Moreover, to optimize makespan and reliability for workflow applications, most existing works use list heuristics rather than genetic algorithms (GAs) which can usually give better solutions. Hence, in this paper, we propose the reliability-driven (RD) reputation, which is time dependent and can be used to evaluate a task's reliability directly using the exponential failure model. Based on RD reputation, we propose a look-ahead genetic algorithm (LAGA) to optimize both time and reliability for a workflow application. LAGA uses a novel evolution and evaluation mechanism: the evolution operators evolve the task-resource mapping for a scheduling solution, while the solution's task order is determined in the evaluation step using our proposed max-min strategy, which is the first two phase strategy that can work with GAs. The experiments show that the RD reputation can improve the reliability for an application with more accurate reputation, while LAGA can provide better solutions than existing list heuristics and evolve to better solutions more quickly than a traditional genetic algorithm.