Private linear regression: Can we scale up with Big Data?

A protocol for secure computation allows two parties with respective private inputs x and y, to collaboratively perform some computation f(x,y) = (f1(x,y), f2(x,y)) such that the first party receives only f1(x,y) and the second party receives only f2(x,y). In this work, we explore the scalability in both number of observations and number of explanatory variables for vertically distributed datasets. The empirical application takes simulated data and carries out benchmark measurement of a suitable extension of the Obliv-c protocol for performing a multivariate regression in an oblivious fashion by harnessing open source libraries. Preliminary results are in single precisions and seem to scale reasonably well with size. The introduction of double precision in the Obliv-c library could be a very boosting feature. Secure Multiparty Regression: Are we ready to keep data privacy promise? Giuseppe Bruno∗ ∗ Bank of Italy, Economics and Statistics Directorate. Abstract Quite often in our societies data on citizens are split up and held by different institutions. For examples we have Social security, Internal Revenues Services, Central Banks and real estate registry. Assuming it is possible to build a single combined databaes collecting all of the information from the individuals present in the different databases coming from different institutions. Many techniques for secure or private computation relies on executing programs in a data-oblivious way, where the same instructions execute independent of the private inputs which are kept in encrypted form throughout the computation. Designers of such computations today must either put substantial effort into constructing a circuit representation of their algorithm, or use a high-level language and lose the opportunity to make important optimizations or experiment with protocol variations. We gauge here the performance and software modification costs for implementing the data oblivious computation. We show a simple extension of the Obliv-C language which allows application developers to program secure computations without being experts in cryptography. Here we focus on methods for computing multivariate linear regression as well as goodness of fit statistics index.Quite often in our societies data on citizens are split up and held by different institutions. For examples we have Social security, Internal Revenues Services, Central Banks and real estate registry. Assuming it is possible to build a single combined databaes collecting all of the information from the individuals present in the different databases coming from different institutions. Many techniques for secure or private computation relies on executing programs in a data-oblivious way, where the same instructions execute independent of the private inputs which are kept in encrypted form throughout the computation. Designers of such computations today must either put substantial effort into constructing a circuit representation of their algorithm, or use a high-level language and lose the opportunity to make important optimizations or experiment with protocol variations. We gauge here the performance and software modification costs for implementing the data oblivious computation. We show a simple extension of the Obliv-C language which allows application developers to program secure computations without being experts in cryptography. Here we focus on methods for computing multivariate linear regression as well as goodness of fit statistics index. JEL classification: C83, D84, E32.