pCloud: A Distributed System for Practical PIR

Computational Private Information Retrieval (cPIR) protocols allow a client to retrieve one bit from a database, without the server inferring any information about the queried bit. These protocols are too costly in practice because they invoke complex arithmetic operations for every bit of the database. In this paper, we present pCloud, a distributed system that constitutes the first attempt toward practical cPIR. Our approach assumes a disk-based architecture that retrieves one page with a single query. Using a striping technique, we distribute the database to a number of cooperative peers, and leverage their computational resources to process cPIR queries in parallel. We implemented pCloud on the PlanetLab network, and experimented extensively with several system parameters. Our results indicate that pCloud reduces considerably the query response time compared to the traditional client/server model, and has a very low communication overhead. Additionally, it scales well with an increasing number of peers, achieving a linear speedup.