A Private-Identifier System From Scalable Private Set Union

Abstract

Modern data-driven collaborations, such as spanning healthcare analysis and advertising attribution, need to merge large identifier sets without revealing which records overlap. A private set union (PSU) protocol enables two parties, each holding a private set X and Y, to securely compute their union X ∪ Y without revealing any additional information. In this work, we revisit recent constructions of PSU protocol to understand it in modular fashion, which enables a better instantiation of such submodules. Furthermore, we further propose an optimization that reduces communication cost, which comes from a novel application permutation-based hashing (phasing). The proposed PSU protocol shows up to 6.9x speed up over high-speed network, and up to 43% reduction in communication costs than the original protocol due for n = 2¹² − 2²⁰ sized set, which is the best performance over LAN network; up to 2.3x faster than the state-of-the-art PSU protocol. Based on this core primitive, we propose a database anonymization system called private-ID (PID). End-to-end evaluations show our PID inherits the PSU improvements, outperforming previous works over high-bandwidth links while remaining competitive on WAN network. Taken together, our results provide a drop-in, high-throughput building block for privacy-preserving data-driven collaborations.