Secure Personalization: Building Trustworthy Recommender Systems

本研究的目的是探索推荐和个性化系统在面对恶意攻击时的漏洞，探索增强其鲁棒性的技术，并研究识别和可能击败攻击的方法。计算机安全的大多数研究都集中在保护组织安全范围内的资产免受未经授权的访问和修改。这个项目检查了被设计为供公众访问和修改的系统的安全问题。我们如何保护这样一个系统免受攻击者试图破坏其功能的合法但有偏见的输入？该项目将增进我们对推荐系统可靠性的理解，目前推荐系统已成为从电子商务、电子学习到内容管理系统等许多领域的重要组成部分。我们将探索针对推荐系统的可能攻击的范围，并开发表征这些攻击及其影响的正式模型。我们将研究评估推荐算法鲁棒性的不同指标，包括准确性、稳定性和对攻击者的预期回报。与此理论工作相结合，我们将使用来自各个领域的数据进行实证调查。我们将测试一系列推荐算法，包括基于用户的、基于项目的和基于模型的协同推荐，并将基于内容和基于知识的协同推荐技术结合起来，探索混合推荐。最后，根据这些结果，我们将考虑如何通过改进算法以及通过检测攻击和适当响应来保护推荐系统。我们的研究将对依赖用户输入来学习用户或组概况的各种自适应信息系统具有重要意义。许多这样的系统都有开放的组件，恶意用户或自动代理可以通过这些组件影响整个系统行为。

The purpose of this research is to explore the vulnerabilities of recommendation and personalization systems in the face of malicious attacks, explore techniques for enhancing their robustness, and examine methods by which attacks can be recognized and possibly defeated. Most research in computer security focuses on protecting assets inside an organization's security perimeter from unauthorized access and modification. This project examines the problem of security for systems that are designed to be accessed and modified by the general public. How do we protect such a system from the legal but biased inputs of an attacker trying to subvert its functionality? The project will advance our understanding of the trustworthiness of recommender systems, now a crucial component in many areas from e-commerce and e-learning to content management systems. We will explore the spectrum of possible attacks against recommendation systems, and develop formal models characterizing these attacks and their impacts. We will investigate different metrics for assessing the robustness of recommendation algorithms including accuracy, stability and expected payoff to the attacker. In tandem with this theoretical work, we will conduct empirical investigations using data from a variety of domains. We will test a range of recommendation algorithms including user-based, item-based and model-based collaborative recommenders, and also explore hybrid recommendation by combining collaborative recommendation techniques with content-based and knowledge-based ones. Finally, informed by these results, we will consider how recommender systems can be secured, through improved algorithms but also by detecting attacks and responding appropriately. Our research will have significant implications for a variety of adaptive information systems that rely on users' input for learning user or group profiles. Many such systems have open components through which a malicious user or an automated agent can affect the overall system behavior.