基于声望的信任管理关键技术研究
文献类型：学位论文
| 作者 | 单明辉 |
| 学位类别 | 博士 |
| 答辩日期 | 2008-06-03 |
| 授予单位 | 中国科学院声学研究所 |
| 授予地点 | 声学研究所 |
| 关键词 | 信任管理 凭证 声望 Beta声望系统 信任融合 不实评价过滤 信任传递 |
| 其他题名 | Research on Key Technology of Reputation-based Trust Management |
| 学位专业 | 信号与信息处理 |
| 中文摘要 | 随着互联网的普及和迅速发展，网络上各种类型的资源和服务也呈爆炸增长趋势。当同类不同质的资源和服务大量存在时，如何从中选择安全、有效、质优的一个，成为用户面临的主要问题之一。信任管理技术是解决该问题的有效手段，它通过在交互的双方中建立合理的信任度，帮助服务请求者在选择前正确评估服务提供者，一方面减少服务请求者权益受侵害的机会，提高了服务请求者的福利，另一方面也促进了服务提供者的优胜劣汰，从而促进网络服务整体环境的繁荣发展。信任管理，按其技术分类可分为基于凭证（策略）的信任管理和基于声望（信誉）的信任管理；前者采用精确的、静态的方式描述处理信任，而后者通过不断的证据收集和信任更新，用一种相对的方法对安全信息进行度量和评估，更适合于处理网络服务环境中受多种动态信息影响的信任关系。 本文研究了以声望为基础的信任管理技术，包括信任的度量、不实评价过滤、信任传递、凭证式和声望式信任的融合等。论文的主要贡献和创新如下： １） 提出一种在信任衰减计算中遗忘因子的动态选择方法以降低信任度计算的总误差。信任会随时间衰减，一般采用对过去事情的重要性“遗忘”的方法来增加近期发生事件的权重。采用较大的衰减因子固然可以使信任度的计算值在真实信任度发生剧烈改变时较快地反映该变化，但同样会由于对过去事件的“遗忘”太多而减少了实际数据量，增加了随机误差。本文将信任度的计算值视为一个随机过程，并以相隔一段时间的两个信任度计算值的差的变化范围是否超出其随机误差，来判定该信任度真实值是否发生了剧烈变化；以此为依据，动态地选择遗忘因子，使得当真实信任度没有剧烈变化时，选用较小的遗忘因子以减小随机误差，当真实信任度发生剧烈变化时，选用较大的遗忘因子以减小跟随误差，从而使信任度计算的总误差比采用单一遗忘因子的方法大大降低。 ２） 提出一种用对数权重均值和方差进行内生式不实评价过滤的方法。该方法的依据是：为尽量大地影响服务提供者的信任度计算值，不实评价与评价总体的偏离一般较大。该方法对交易次数设置一分位值上限，以此限制虚报过大交易量的不实评价；以交易次数的对数做为权重进行信任度均值计算，以此增加对小交易量评价者意见的考虑；以加权方差定义容限，识别并过滤与信任度平均值差别较大的评价。该方法能有效过滤不实评价，提高声望计算的精度，并且不需要评价者信息，因此尤其适用于匿名评价的环境下。 ３） 提出一种以平均偏离度为基础的外生式不实评价过滤的方法。该方法以服务请求者之间在历史评价活动中体现出来的一般偏离程度为基础，定义服务请求者之间的平均偏离度。在其后收到某服务请求者的评价时，则依据其平均偏离度对其评价的权重做衰减，以使得平均偏离度大的服务请求者，其评价的权重衰减也大，以此实现不实评价的过滤。该方法对不实评价过滤效果显著，同时对正常评价的误识率也非常低。 ４） 提出一种借助三角不等式进行平均偏离度计算的信任传递方法，以增加可用的评价量，提高了计算精度。该方法以平均偏离度为距离度量，对任意两个服务请求者之间的路径，以三角不等式进行距离界计算，在不同的假设下以不同界作为其平均偏离度；以平均偏离度矩阵为基础，给出计算全局偏离度矩阵的方法；为降低负载，给出了一种近似计算方法，适用于在集中式信任管理环境。该信任传递方法大大增加了可用评价的来源范围，使得服务请求者可以对任意一个服务提供者进行信任计算，提高了计算的精度，拓宽了信任管理的应用环境。 ５） 提出了一种融合凭证和声望的信任管理模型。该模型以声望式信任管理系统为基础，在凭证学习阶段通过凭证与声望之间的相关性，动态自适应地学习凭证的作用；而在凭证使用阶段，通过学习阶段积累的知识，将凭证转化为一定的声望值；在综合阶段，将由凭证转化的声望值与声望系统输出的声望值综合，输出最终信任度。该融合方法使得信任管理系统可以接受评价和凭证两种类型的数据输入，既具有声望式信任管理系统可以适应动态、开放、不确定信任环境的优点，又在评价数据不足时可以依据凭证导出一定的信任度，供服务使用者参考。相比传统方法，本方法可大幅提高信任系统的性能，在同样的数据量情况下可以提高计算信任的精度，在同样的计算精度要求下降低对数据量的需求，尤其适用于服务提供者的交互历史较少的情况。 |
| 英文摘要 | Along with the popularization and development of Internet, the number of resources and services increases exponentially. When resources or services of the same type but with different qualities are under consideration, it remains the main problem that how to choose a secure, effective and high-quality resource or service among them. Trust management technology is an effective way to solve this problem. It can set up a reasonable trust degree between the two side of interaction, and help the service requester to evaluate the service provider correctly before they conduct a real transaction. By doing so, trust management system can on one hand reduce the possibility for the service requesters of being harmed, and raise their welfares. On the other hand, the system can accelerate the process of letting bad providers down, promoting the entire environment of Internet online service. Trust management can be classified into two types: Credential-based and reputation-based; the former describes trust as precise and static, while the latter employs a relative method to measure and evaluate the secure information when collecting evidence and updating trust value continuously, so reputation-based trust management is more effective for processing trust relationship affected by many dynamic factors in Internet online service environment. This paper studies reputation-based trust management technologies, including trust measurement, unfair rating filtering, trust transfer, and combination of reputation and credential. The main contributions are as follows: 1) A method of dynamicly choosing forgetting factor in trust decaying is proposed. Trust decays with time, so evidence is weighted according to the passed time. Big forgetting factor can reflect change of real trust estimation value quickly, while may decrease the evidence amount because of “over-forgeting” of the older evidence, and thus cause the augment of random error. This paper considers the computed trust value as a stochastic process, and judges a change of real trust value by watching if the difference of two computed trust value is bigger than its random error. A small forgetting factor is used when a big change of real trust value change is detected; otherwise a big forgetting factor is applied. The simulation results show that the total error is reduced greatly compared with the method using fixed forgetting factor. 2) An endogenous method of unfair rating filering is proposed based on logarithmic weights and mean values. The base of the method is that an unfair rating is usually apart from the average rating a lot in order to be influential. In the method a place value is set as the interaction number’s limit, in order to limit the bragging of interaction; the average of trust degree is calculated with weights of logarithms of interaction numbers, to augment the weight of ratings from small dealer; a threshold is defined as weighted variance, to detect and filter out ratings far away from the average trust degree. This method can filter out unfair ratings effectively and improve the accuracy of trust calculation. Moreover, no requester information is needed, so it’s especially fit for anonymous environments. 3) An extraordinary method of unfair rating filtering is proposed based on average departure degree. The method employs service requester’s normal departure degree demonstrated in the past rating activities. On receiving a rating from a requester, the weight of rating is attenuated according to the requester’s average departure degree, in order to lower down the weights of ratings from requester whose departure degree is high, so to filter out the unfair ratings. In simulations it’s showed that the effectiveness of this method is notable, and the false positive rate is very low, too. 4) A trust transfer method is proposed based on triangle inequation of average departure degree. The method employs average departure degree as metric of distance between two service requesters, and uses triangle inequation to calculate the bounds of the distance, then corresponding bound is used as their average departure degree under different circumstances. Based on average departure degree matrix, a algorithm for calculating the global average departure degree matrix is proposed. And for centralized system, an approximate algorithm is presented to lower down the overload. The trust transfer method enlarges the scope of trust sources, and enables a requester to calculate any provider’s trust degree, besides makes the calculation more accurate, and enlarges the application domains of trust management systems. 5) A trust management model of combining credential and reputation is presented. The model is based on reputation-based trust management. At credential learning stage, through the correlation between reputation and credential, the model can learn the credential adaptively; at credential working stage, the credential is transformed into some reputation with help of the results learned at the credential learning stage; and at synthesis stage, the reputations transformed from credentials and ratings are melted together to output the final trust degree. This combining method can enable the trust management to accept inputs of both ratings and credentials, inheriting the merits of adapting to dynamic, open, uncertain trust environments from reputation-based trust management system, besides having the ability of exporting trust degree when no rating is available. Compared with the traditional trust management system, this model can improve the system performance evidently, which means increasing the accuracy of calculation with the same data, or lowering down the demand of data amount with the same accuracy requirement. This method is especially fit for circumstances in which the provider’s interactions are few. |
| 语种 | 中文 |
| 公开日期 | 2011-05-07 |
| 页码 | 135 |
| 源URL | [http://159.226.59.140/handle/311008/332]  |
| 专题 | 声学研究所_声学所博硕士学位论文_1981-2009博硕士学位论文 |
| 推荐引用方式 GB/T 7714 | 单明辉. 基于声望的信任管理关键技术研究[D]. 声学研究所. 中国科学院声学研究所. 2008. |
入库方式： OAI收割
来源：声学研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。