Towards reliable reputations for distributed applications.

摘要

The Internet today is experiencing a rapid growth in popularity of online communities like P2P file sharing networks, Wikis, and distributed web applications like Flickr and YouTube. The open and anonymous nature that makes interacting in these applications so popular also makes them vulnerable to attacks from malicious and self-interested users. A large amount of research proves that digital reputations are an effective solution to the problem of user maliciousness in such networks. A reputation is a statistical estimate of a user's trustworthiness computed from feedback given by previous transaction partners to the user. There are some fundamental "reliability" concerns, however, that limit the effectiveness of existing reputations mechanisms and their potential impact in network protocols and distributed applications today. Current reputation schemes are still highly vulnerable to false ratings attacks posed by malicious individuals, Sybils, and colluders. Furthermore, high churn rates in dynamic networks pose a serious challenge to the adoption of reputation systems that depend on long-term participation for accuracy.We believe that, when building distributed applications, trust should be as important a consideration as traditional issues like latency and fault-tolerance. In this dissertation, therefore, we address the critical reliability concerns that limit the effectiveness of reputations. We develop a scalable and globally decoupled reputations approach that disassociates every user's service provider reputation from its service referral reputation, thereby making reputations robust to feedback tampering. We also design a statistical metric that estimates reputation inflation via collusion, and demonstrate its effectiveness on eBay.com reputation profiles. To address the challenge posed by network churn, we develop proactive reputations, an approach that establishes quick and accurate reputations for unknown peers or newcomers in reputation systems. Finally, this dissertation compares feedback-based reputations with a new and increasingly popular paradigm to establish digital trust, one presented by social networks.In summary, this dissertation significantly improves the state-of-the-art in reputation systems and advances the growth of next-generation distributed applications by improving their resistance to trust-based vulnerabilities.