【作者】 冯玉翔；

【导师】 唐韶华；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2013， 博士

【摘要】 随着Internet技术的发展，大规模分布式计算环境，如P2P计算、网格计算、AdHoc网络、社会网络等得到了广泛的应用。这类计算环境下的新的应用模式通常表现为由多个异构的系统或软件服务组成的动态协作系统。这类系统的形态正从封闭的、面向熟识用户群体的、相对静态的形式，向开放的、面向公共用户的、动态协作的模式转变。这些特性使得这类系统表现出很强的复杂性和动态性，也使得我们很难通过“控制”来适应这种环境，即很难用传统的中央控制机制来管理这类系统的安全。同时，以P2P文件分享系统、大型电子商务系统、社交网站为代表的新的应用模式也带来了信息过载、节点自私性等新的管理问题。现实中人类社会也具备上述的特点并且更为显著。信任是人类社会的重要基石之一。在社会科学、商业等诸多领域中，特别是人们的日常生活中，信任都在时刻发生着作用。所以，近年来计算机科学领域中的信任管理模型的研究主要是通过借鉴社会学、心理学等领域中有关信任的研究成果，通过“信任”来简化上述系统的复杂性。本文主要围绕上述的大规模分布式计算环境的特点，研究适合这些特点的动态信任管理机制来解决此环境下的应用的安全管理问题，并在此基础上探索了利用动态信任管理机制来解决此环境下新的应用模式带来的信息过载、节点自私性问题的可能方法。本文的主要研究内容和创新点归纳如下：1）目前的信任管理机制的研究中，通常采用数学以及人工智能等方面的理论和方法来形式化的描述一定环境下的信任关系，从而实现此环境下的信任评估。本文首先分析了大规模分布式计算环境下信任和信任管理的特点，借鉴了信任在其他学科的研究成果，归纳总结了计算机科学领域中信任研究的进展情况。在此基础上，本文定义了目前有代表性的动态信任关系的概念和动态信任关系的基本性质，给出了适应大规模分布式计算环境需求的动态信任管理模型所必须具备的特征，讨论了信任的度量方法。综合上述研究内容，本文设计了一个基于声誉的分布式动态信任管理模型。该模型能够成为传统安全技术的一种有益补充，作为“软安全”机制，与传统的安全机制协作实现大规模分布式计算环境下的系统安全性。2）现有的大多数信任管理模型都是使用固定的公式和预定义的参数等来表述信任关系。它们常常面临两个问题：一是大规模分布式计算环境下的极端动态性导致这些固定的公式和预定义的参数很难适用。二是如果要建立的信任关系必须面向信任的多个属性，由于多属性之间的关联常常是模糊、不确定的，而固定的公式和预定义的参数很难描述这种模糊、不确定性。本文基于进化计算中的“涌现”性的学习模型，将机器学习引入信任管理，提出了一个针对多属性信任关系的动态评估模型。该模型能综合对信任的多属性进行评估，并能够在不断的执行动态评估任务的过程中，自主的学习，自主的提高模型的评估任务的执行能力，从而适应上述的动态性和模糊性。3）解决新的应用模式下存在的信息过载问题，在海量的数据中迅速准确的找到用户可能感兴趣的内容并推荐给用户，对于今天的大型电子商务应用系统来说是一个至关重要的问题。以协同过滤为代表的推荐技术是当前的一个研究热点。我们探索了动态信任管理在个性化推荐系统的应用方法。我们将动态信任管理模型和协同过滤技术结合起来，设计了一个基于信任的个性化推荐系统。该系统将信任度、评价度和推荐度统一起来，利用用户间的评价相似性生成信任关系，并用前述的信任模型来管理这些信任关系。这样，用户就可以根据信任关系及其他用户给出的评价来预测自身对未知资源目标的评分。从而利用信任管理来解决协同过滤中由于用户评价项目的稀疏性、评价信息不准确等问题带来的推荐准确度不高的问题。4）合作涌现问题是指“自私”的个体群体中为何会出现“合作”行为，它是多个学科领域的研究热点。这个问题在计算机科学中最重要的应用是：在以P2P文件共享等为代表的应用中，系统由异构的个体构成，这些个体必须通过合作来获得系统的总体公共目标。因而个体的“自私性”必将导致系统中free-riding问题的出现。我们认为信任关系在人类社会的合作涌现中起着非常重要的作用。因此在本文中对以信任为基础的合作激励机制进行了探索。我们设计了一个基于经典囚徒困境问题的演化游戏。在游戏中我们加入了以信任为基础的合作激励机制，即交互的双方的行为除了会影响到双方的收益之外，还会影响到双方的声誉变化。而其他的交互者则会根据对方的声誉来选择自身的行为。演化实验的结果说明，信任在游戏中成功的促进了合作，起到了合作激励的作用。更多还原

【Abstract】 With the development of Internet technology, large-scale distributed computing environ-ments such as P2P computing, grid computing, Ad Hoc networks, social networks have beenwidely used. The new application models in such computing environment typically be com-posed by multiple heterogeneous systems or software services. Now, such models moving fromthe closed, facing familiar user groups, relatively static form, to open for public users, dynamiccollaboration paradigm shift. These features give the great complexity and dynamic to thesesystems. It is impossible to "Control" this systems, i.e. that is difficult to use traditional cen-tral control mechanisms to manage the security of such systems. Meanwhile, P2P file sharingsystem, large e-commerce systems, social networking sites as a representative of these new ap-plication models also brings information overload, the "selfish" node and other new issues.Human society also has the above characteristics, and maybe more significant."Trust"is one of the cornerstones of human society. In the social sciences, business and many otherareas, especially in people’s daily life, trust is happening in the moment of action. Therefore,by drawing sociology, psychology and other fields of research about trust, trust management iswidely researched in computer science in recent years. These researches want to use "trust" tosimplify the complexity of the systems.Thisthesismainlyfocusonthecharacteristicsoflarge-scaledistributedcomputingenviron-ment. We has proposed suitable dynamic trust management mechanisms for this environment,and explored the possible ways to address the new issues such as the information overload and"selfish" node by using trust management mechanisms. The main contents and novelties areillustrated as follows:1)The current trust management mechanisms, usually using mathematical, artificial intelli-gence and other aspects of the theory and methods to describe the formal relations of trust undercertain circumstances, in order to achieve trust evaluation in these environment. We analyzedthe characteristics of dynamic trust in large-scale distributed computing environment by draw-ing sociology, psychology and other fields of research about trust, summarized the progressof the research progress of trust in computer science, discussed the characteristics of dynamictrust management mechanism in large-scale distributed computing environment. On this ba- sis, we proposed a reputation-based distributed dynamic trust management mechanism. Thismechanism, as a "soft security" mechanism, can be a useful supplement to the traditional secu-rity mechanism. This mechanism can collaborate with the traditional security mechanisms toachieve security in large-scale distributed computing environment.2)Most of the existing trust management mechanisms are using fixed formulas and prede-fined parameters to describe the trust relationship. These mechanisms always have two prob-lems: one is the fixed formulas and predefined parameters can not adopt the dynamic envi-ronment; the other is the fixed formulas and predefined parameters can not describe the fuzzyrelationship between different attributes of trust. We designed a dynamic evaluation model formulti-attributes trust relationship. This model achieves the learning ability by using evolution-ary computing. With autonomous learning, the model can improve it’s evaluation performanceof its own. Therefore, this model can adopt the the dynamic environment and describe the fuzzyrelationship between different attributes of trust.3)Addressing the information overload is a crucial issue for today’s large-scale e-businessapplications. Collaborative filtering is a hot-spot in recent research of personalized recommen-dation systems. We argued the relationship between collaborative filtering and trust manage-ment. Byunifyingthetrust,theevaluation,andtherecommendation,wedesignedapersonalizedrecommendation systems based on reputation. User can predict his rating on the unknown re-sources by the reputation of the other users and the recommendation of the others. By using thetrust management mechanism, the system can address the problems caused by sparse rating andunreliable rating. Therefore, the system can improve the accuracy of the prediction.4)The evolution of cooperation between selfish individuals is a fascinating topic that hasattracted scientists in different fields. This question is also beginning to arise in computer sys-tems. The P2P systems are characterized by heterogeneous components that interact to achievethe public goals. Thus, the free-riding problem will caused by selfish individuals in these sys-tems. We argued that the trust information is an important factor in the rise of cooperativebehaviors in the evolution of a group of selfish individuals. Therefore, we explored the cooper-ation stimulation mechanisms based on trust. This thesis designed an evolutionary game basedon classic prisoners dilemma game. A cooperation stimulation mechanism based on trust hasbeen put in this game. The individuals update their strategies and adjust their relationship with others merely according to the opponents’ trust information. The result of the game illustratedthe effectiveness of our mechanism.更多还原