Private computation on common records refers to analyze data from two databases containing shared records without revealing personal information. As a basic requirement for private computation, the databases involved essentially need to be aligned by a common identification system. However, it is hard to expect such common identifiers in real world scenario. For this reason, multiple quasi-identifiers can be used to identify common records. As some quasi-identifiers might be missing or have typos, it is important to support fuzzy records setting. Identifying common records using quasi-identifiers requires manipulation of highly sensitive information, which could be privacy concerns. This work studies the problem of enabling such data analysis on the fuzzy records of quasi-identifiers. To this end, we propose "ordered threshold-one (OTO)" matching which can be efficiently realized by circuit-based private set intersection~(CPSI) protocols and some multiparty computation (MPC) techniques. Furthermore, we introduce some generic encoding techniques from traditional matching rules to the OTO matching. Finally, we achieve a secure efficient private computation protocol which supports various matching rules which have already been widely used. We also demonstrate the superiority of our proposal with experimental validation. First, we empirically check that our encoding to OTO matching does not affect accuracy a lot for the benchmark datasets found in the fuzzy record matching literature. Second, we implement our protocol and achieve significantly faster performance at the cost of communication overhead compared to previous privacy-preserving record linkage (PPRL) protocols. In the case of 100K records for each dataset, our work shows 147.58MB communication cost, 10.71s setup time, and 1.97s online time, which is 7.78 times faster compared to the previous work (50.12 times faster when considering online time only).

Machine learning on (homomorphic) encrypted data is a cryptographic method for analyzing private and/or sensitive data while keeping privacy. In the training phase, it takes as input an encrypted training data and outputs an encrypted model without ever decrypting. In the prediction phase, it uses the encrypted model to predict results on new encrypted data. In each phase, no decryption key is needed, and thus the data privacy is ultimately guaranteed. It has many applications in various areas such as finance, education, genomics, and medical field that have sensitive private data. While several studies have been reported on the prediction phase, few studies have been conducted on the training phase.In this paper, we present an efficient algorithm for logistic regression on homomorphic encrypted data, and evaluate our algorithm on real financial data consisting of 422,108 samples over 200 features. Our experiment shows that an encrypted model with a sufficient Kolmogorov Smirnow statistic value can be obtained in ∼17 hours in a single machine. We also evaluate our algorithm on the public MNIST dataset, and it takes ∼2 hours to learn an encrypted model with 96.4% accuracy. Considering the inefficiency of homomorphic encryption, our result is encouraging and demonstrates the practical feasibility of the logistic regression training on large encrypted data, for the first time to the best of our knowledge.

In this paper, we describe a polynomial time cryptanalysis of the (approximate) multilinear map proposed by Coron, Lepoint, and Tibouchi in Crypto13 (CLT13). This scheme includes a zero-testing functionality that determines whether the message of a given encoding is zero or not. This functionality is useful for designing several of its applications, but it leaks unexpected values, such as linear combinations of the secret elements. By collecting the outputs of the zero-testing algorithm, we construct a matrix containing the hidden information as eigenvalues, and then recover all the secret elements of the CLT13 scheme via diagonalization of the matrix. In addition, we provide polynomial time algorithms to directly break the security assumptions of many applications based on the CLT13 scheme. These algorithms include solving subgroup membership, decision linear, and graded external Diffie–Hellman problems. These algorithms mainly rely on the computation of the determinants of the matrices and their greatest common divisor, instead of performing their diagonalization.