Secure and Accurate Two-Step Hash Encoding for Privacy-Preserving Record Linkage

Abstract

In order to discover new insights from data, there is a growing need to share information that is distributed across multiple databases that are often held by different organisations. One key task in data integration is the calculation of similarities between records to identify pairs or sets of records that correspond to the same real-world entities. Due to privacy and confidentiality concerns, however, the owners of sensitive databases are often not allowed or willing to exchange or share their data with other organisations to allow such similarity calculations. In this paper we propose a novel privacy-preserving encoding technique that can be used to securely calculate similarities between sensitive values held in different databases. Our technique uses two-step hashing to encode values into an integer set representation that provides strong privacy guarantees and allows accurate similarity calculations. We provide a theoretical analysis of the accuracy and privacy of our encoding technique, and conduct an empirical study on large real databases containing several millions records. Our results show that our technique provides high security against privacy attacks and achieves better similarity accuracy compared to two state-of-the-art encoding techniques.

A Graph Matching Attack on Privacy-Preserving Record Linkage

“…Step Hash Encoding (2SH). This recently proposed novel PPRL encoding method [27] aims to address both privacy issues of BF encoding as well as the high computational costs of TMH encoding. This method employs two efficient hashing steps based on bit vectors and an integer mapping to provide accurate Jaccard similarity calculations with improved privacy protection.…”

“…( 1 , 2 ), where () is the function used to calculate the similarity between two nodes. This can, for example, be the Dice or Jaccard similarity depending on the encoding method used [27,30,34].…”

“…Using the obtained min-hash signatures, bit arrays of length = 1000 were generated for each encoded value [34]. For 2SH encoding, we used the parameter settings: = 2, = 1000, and = [10,15,20] [27]. The accuracy results discussed below for each database were averaged over all these parameter settings.…”

“…To overcome the privacy weaknesses of BF encoding, alternative perturbation based encoding techniques for PPRL have been developed. One of these uses tabulation Min-hashing [34] while another recently proposed method uses two hash encoding steps [27]. We discuss these two techniques in detail in Sect.…”

“…To provide the reader with an understanding of encoding methods used for PPRL, we now briefly describe the three methods we will use in the experimental evaluation of our attack: popular and practically used Bloom filter (BF) encoding [3,6,26,28,30], and the more recently proposed tabulation min-hashing (TMH) [34] and two-step hashing (2SH) [27] methods. We then provide an overview of existing attack methods on PPRL in Sect.…”