Privacy Preserving Joins

In this paper, we design a system for mutually distrustful entities to perform privacy preserving joins, leveraging the power of a memory-limited secure coprocessor. Under this setting, we critique a questionable assumption in a previous privacy definition [1] that leads to unnecessary information leakage. We then remove the assumption and propose a new definition. Based on this definition, we propose three correct and provable secure algorithms to compute general joins of arbitrary predicates, by utilizing available cryptographic tools in a nontrivial way. We discuss different memory requirements of our proposed algorithms, and explore how to trade little privacy with significant performance improvement. In [2], we evaluate the performance of our algorithms by numerical examples. We also show the performance superiority of our approach over secure multi-party computation in [2].