Secure bi-attribute index: Batch membership tests over the non-sensitive attribute

Abstract

Secure index techniques enable keyword searches on encrypted univariate data, but they struggle with bi-attribute data common in AI and data mining applications. Traditional approaches suffer from inefficiencies during prefix queries due to duplicate trapdoor generations. Although plaintext processing of one non-sensitive attribute can boost performance, it may also introduce privacy risks from inter-attribute correlation and potential inference attacks. This paper presents a secure bi-attribute indexing solution, illustrated with a case study on searchable encryption for time-series data. We introduce two variants of matrix Bloom filters tailored for different workloads and implement a concept of bounded privacy loss via noise infusion from the randomized response technique. The outcome adheres to locally differential privacy principles, offering a provable privacy guarantee for sensitive attribute items.