On the use of decentralization to enable privacy in web-scale recommendation services

Abstract

We present the design, implementation, and evaluation of a decentralized framework for enabling privacy in Web-scale recommendation services. Our framework, which comprises of a decentralized middleware that is hosted and run by federated entities, is designed to support collaborative-filtering and content-based recommendations. We design a novel distributed protocol that clusters users into interest groups comprised of like-minded members and ensures a desired minimum size (k-anonymity parameter), while keeping user profiles on client-side only. In order to aggregate users' consumption for the purpose of generating recommendations, we design a novel decentralized aggregation mechanism that protects against auxiliary information attacks that have crippled conventional k-anonymity based systems. Our prototype system ensures that the desired k-anonymity level is met, and can prevent auxiliary information attacks using a middleware of modest size, and is empirically shown to be resistant to moderate degree of collusion. While preserving privacy, our system enables effective clustering of like-minded users, and offers good quality of recommendations. Also, the prototype's decentralized design and lightweight protocols enable almost linear-scaling with increased size of the middleware.