Provably Secure Decisions Based on Potentially Malicious Trust Ratings

任何使用互联网的人都会注意到评分和评论，例如在预订酒店时。我们能在多大程度上信任我们在网上看到的评论？也许是网络喷子用负面评价轰炸一个网站，或者是一家公司的公关人员为他们的客户写了一些非常积极的东西？大多数人都有一定程度的怀疑。评级也可以在屏幕后面使用，例如在标记可能的恶意软件时。我们能基于评分自动做出决策吗？有没有一种公式化的方法来使用评级而不被欺骗？我们的研究为基于评级的安全决策提供了一个基础。在开放网络中，收视率尤其重要，因为开放网络在物联网中扮演着重要角色。在开放网络中，参与者是潜在的恶意（攻击者），但我们可能依赖于他们提供的信息。在当前对使用可能不公平评级的网络的分析中，对攻击者进行了假设。例如，他们最大化他们的利润，或者想要执行特定的行动。然而，在现实中，我们无法知道攻击者想要或将要做什么。这是我们方法中的关键挑战：无论攻击者的行为如何，我们都提供具有经过验证的风险界限的解决方案。利用信息理论，数字网络能够在有噪声的情况下重建信号。通过将正确的决策建模为信号，将攻击建模为噪声，我们在之前的工作中已经证明，通常情况下，信息是通过评级来传递的。使用类似于数字网络中应用的技术，我们可以重建正确的决策。因此，我们提出了一个方法框架，利用信息来做出正确的决策。我们的框架包括将评级转换为正确决策的一般技术，以及基于这些技术的决策方案。有两个主要的应用：一个是做决定的中央系统，另一个是个人做决定的分散系统。集中的例子包括YouTube基于版权要求删除内容，Facebook审查淫秽内容，以及在电子商务系统中发现欺诈商家。去中心化的例子是ad-hoc网络，其中远程节点被选择路由敏感信息，点对点网络，恶意的对等破坏协议，以及对等评估，其中学生必须给他们的同伴打分。我们提供了一个集中和分散的系统，可以在所有攻击下做出可证明的正确决策。该框架的一个主要组成部分是评级的理论基础。我们定义了三个理想的性质：鲁棒性、最优性和稳定性。如果决策方案提供的错误决策的概率小于等于，则该决策方案被称为-鲁棒决策方案。从足够诚实的来源获得足够的评级，这很容易获得。最优性是关于将成本（评级的数量和复杂性）降低到最小。稳定性是指如果诚实度低于预期，在不增加成本的情况下无法改进决策方案。我们将研究在哪些情况下鲁棒性、最优性和稳定性可以结合，以及这种结合的代价是什么。最有趣的上下文是动态的：用户可以确定（通过假阳性/假阴性的概率）先前评级的准确性。这种动态环境在理论上和实践上都很有趣。理论上的兴趣是需要更先进的信息理论技术，并且可能与其他领域有很深的联系，例如对抗性机器学习。实际的利益是，在许多系统中，资源被使用了不止一次，决策者对旧评级的质量确实有一个模糊的概念。可证明的是，这种动态信息的有效利用尚未实现，并将提高评级系统的安全性。这项研究的结果将是提供更安全的评级系统。

Anyone who uses the internet will be aware of ratings and reviews, for example when booking a hotel. How much trust can we place in reviews we read online? Perhaps internet trolls bombarded a site with negative ratings, or perhaps a company's PR person wrote something glowingly positive for their client? Most people have a degree of skepticism. Ratings can also be used behind the screens, for example when flagging possible malware. Can we automate decisions based on ratings? Is there a formulaic way of using the ratings without being deceived? Our research proposes a foundation to enable secure decisions based on ratings.Ratings are especially important in open networks, which play a large role in the internet of things. In open networks, participants are potentially malicious (attackers), yet we may rely on information that they provide. In current analysis of networks that use potentially unfair ratings, assumptions are made about the attackers. For example, that they maximise their profit, or want to perform specific actions. In reality, however, we cannot know what the attackers want or will do. This is the crucial challenge in our approach: we provide solutions with a proven risk-bound, regardless of the behaviour of the attackers.Using information theory, digital networks are able to reconstruct signals despite noise. By modelling correct decisions as a signal, and attacks as noise, we have proven in previous work that typically, information is carried in ratings. With techniques similar to those applied in digital networks, we can reconstruct the correct decision. So, we propose a framework of methods to use information to come to correct decisions despite attacks.Our framework consists of general techniques regarding transforming ratings to correct decisions, and of decision schemes based on these techniques. There are two major applications: a centralised system making a decision, and a decentralised system where individuals make decisions. Centralised examples are YouTube deleting content on the basis of copyright claims, Facebook censoring obscene material and finding fraudulent merchants on an e-commerce system. Decentralised example are ad-hoc networks, where distant nodes are selected to route sensitive information, peer to peer networks, with malicious peers breaking protocol, and peer assessment, where students have to grade their peers. We deliver both a centralised and a decentralised system that makes provably correct decisions under all attacks.A major component of the framework is the theoretical foundation for ratings. We define three desirable properties: robustness, optimality and stability. A decision scheme is called epsilon-robust if it provides the wrong decision with a probability under epsilon. With sufficient ratings from sources that are sufficiently probably honest, this is easy to obtain. Optimality is about reducing the cost (amount and complexity of ratings) to the minimum. Stability means that if the degree of honesty is lower than expected, the decision scheme cannot be improved without raising costs. We investigate in which contexts robustness, optimality and stability can combine, and at which cost this occurs.The most interesting context is dynamic: where users can determine (with a probability of false positives/negatives) the veracity of previous ratings. This dynamic context is both theoretically and practically interesting. The theoretical interest is that more advanced information theoretic techniques are required, and there may be deep links to other fields, such as adversarial machine learning. The practical interest is that in many systems, sources are being used more than once, and decision makers do have a vague idea about the quality of older ratings. Provably effective use of this dynamic information has not been achieved, and will improve the security of rating systems.The result of this research will be to provide more secure rating systems.