Privacy preserving linkage and sharing of sensitive data
Date
2018
Authors
Lazrig, Ibrahim Meftah, author
Ray, Indrakshi, advisor
Ray, Indrajit, advisor
Malaiya, Yashwant, committee member
Vijayasarathy, Leo, committee member
Ong, Toan, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Sensitive data, such as personal and business information, is collected by many service providers nowadays. This data is considered as a rich source of information for research purposes that could benet individuals, researchers and service providers. However, because of the sensitivity of such data, privacy concerns, legislations, and con ict of interests, data holders are reluctant to share their data with others. Data holders typically lter out or obliterate privacy related sensitive information from their data before sharing it, which limits the utility of this data and aects the accuracy of research. Such practice will protect individuals' privacy; however it prevents researchers from linking records belonging to the same individual across dierent sources. This is commonly referred to as record linkage problem by the healthcare industry. In this dissertation, our main focus is on designing and implementing ecient privacy preserving methods that will encourage sensitive information sources to share their data with researchers without compromising the privacy of the clients or aecting the quality of the research data. The proposed solution should be scalable and ecient for real-world deploy- ments and provide good privacy assurance. While this problem has been investigated before, most of the proposed solutions were either considered as partial solutions, not accurate, or impractical, and therefore subject to further improvements. We have identied several issues and limitations in the state of the art solutions and provided a number of contributions that improve upon existing solutions. Our rst contribution is the design of privacy preserving record linkage protocol using semi-trusted third party. The protocol allows a set of data publishers (data holders) who compete with each other, to share sensitive information with subscribers (researchers) while preserving the privacy of their clients and without sharing encryption keys. Our second contribution is the design and implementation of a probabilistic privacy preserving record linkage protocol, that accommodates discrepancies and errors in the data such as typos. This work builds upon the previous work by linking the records that are similar, where the similarity range is formally dened. Our third contribution is a protocol that performs information integration and sharing without third party services. We use garbled circuits secure computation to design and build a system to perform the record linkages between two parties without sharing their data. Our design uses Bloom lters as inputs to the garbled circuits and performs a probabilistic record linkage using the Dice coecient similarity measure. As garbled circuits are known for their expensive computations, we propose new approaches that reduce the computation overhead needed, to achieve a given level of privacy. We built a scalable record linkage system using garbled circuits, that could be deployed in a distributed computation environment like the cloud, and evaluated its security and performance. One of the performance issues for linking large datasets is the amount of secure computation to compare every pair of records across the linked datasets to nd all possible record matches. To reduce the amount of computations a method, known as blocking, is used to lter out as much as possible of the record pairs that will not match, and limit the comparison to a subset of the record pairs (called can- didate pairs) that possibly match. Most of the current blocking methods either require the parties to share blocking keys (called blocks identiers), extracted from the domain of some record attributes (termed blocking variables), or share reference data points to group their records around these points using some similarity measures. Though these methods reduce the computation substantially, they leak too much information about the records within each block. Toward this end, we proposed a novel privacy preserving approximate blocking scheme that allows parties to generate the list of candidate pairs with high accuracy, while protecting the privacy of the records in each block. Our scheme is congurable such that the level of performance and accuracy could be achieved according to the required level of privacy. We analyzed the accuracy and privacy of our scheme, implemented a prototype of the scheme, and experimentally evaluated its accuracy and performance against dierent levels of privacy.
Description
Rights Access
Subject
record-linkage
privacy
security