Privacy Attack on Multiple Dynamic Match-key based Privacy-Preserving Record Linkage

Abstract

Introduction
 Over the last decade, the demand for linking records about people across databases has increased in various domains. Privacy challenges associated with linking sensitive information led to the development of privacy-preserving record linkage techniques. The multiple dynamic match-key encoding approach recently proposed by Randall et al. (IJPDS, 2019) is such a technique aimed at providing sufficient privacy for linkage applications while obtaining high linkage quality. However, the use of this encoding in large databases can reveal frequency information that can allow the re-identification of encoded values.
 Objectives
 We propose a frequency-based attack to evaluate the privacy guarantees of multiple dynamic match-key encoding. We then present two improvements to this match-key encoding approach to prevent such a privacy attack.
 Methods
 The proposed attack analyses the frequency distributions of individual match-keys in order to identify the attributes used for each match-key, where we assume the adversary has access to a plain-text database with similar characteristics as the encoded database. We employ a set of statistical correlation tests to compare the frequency distributions of match-key values between the encoded and plain-text databases. Once the attribute combinations used for match-keys are discovered, we then re-identify encoded sensitive values by utilising a frequency alignment method. Next, we propose two modifications to the match-key encoding; one to alter the original frequency distributions and another to make the frequency distributions uniform. Both will help to prevent frequency-based attacks.
 Results 
 We evaluate our privacy attack using two large real-world databases. The results show that in certain situations the attack can successfully re-identify a set of sensitive values encoded using the multiple dynamic match-key encoding approach. On the databases used in our experiments, the attack is able to re-identify plain-text values with a precision and recall of both up to 98%. Furthermore, we show that our proposed improvements are able to make this attack harder to perform with only a small reduction in linkage quality.
 Conclusions 
 Our proposed privacy attack demonstrates the weaknesses of multiple match-key encoding that should be taken into consideration when linking databases that contain sensitive personal information. Our proposed modifications ensure that the multiple dynamic match-key encoding approach can be used securely while retaining high linkage quality.

A methodological assessment of privacy preserving record linkage using survey and administrative data

“…It should be noted that while we performed an assessment of data quality, we did not conduct a separate privacy analysis of the encoded databases. Prior to a full implementation of PPRL at NCHS, a privacy risk-reidentification analysis would need to be conducted as well [24][25][26]. This would enable us to identify any potential known risks and assess if there are any mitigation strategies that could be implemented to maintain the protection of Nevertheless, the data quality results of this case study are encouraging.…”

Validating a novel deterministic privacy-preserving record linkage between administrative & clinical data: applications in stroke research

“…Another key consideration is the security and privacy offered by different linkage methods. While Bloom filter and multiple match key encoding are more efficient methods of linking, there may be some trade-offs between data linkage quality and vulnerability to privacy attacks [32,[34][35][36]. Homomorphic encryption algorithms offer exceptional security, and the proposed methodology delivers a highly accurate data linkage, with the trade-off instead being computational resources.…”

Privacy-preserving record linkage using autoencoders

“…To quantify the similarity of the distribution of a set of autoencoder encodings compared to a normal distribution N (0, 1), we use the histogram intersection measure [38,45] 1 . For each of the d dimensions, we generate a histogram he of the encodings using b bins of equal width and similarly generate a histogram hn for a normal distribution with the same width and the same number of bins and data points as for the number of encodings.…”