The Privacy Bootstrap

Abstract

Methods for the privacy protection of microdata include grouping, deleting records or adding simulated records, data swapping, and the publication of data perturbed with random noise. We suggest a variant of the latter in which the noise is generated by bootstrapping from the original empirical distribution. The published data distribution then essentially consists of a convolution of a distribution with itself, and the distribution can be recovered, although the individual observations remain protected. By means of a regression example, we explore the trade-off between privacy protection based on bootstrapping and the efficiency of estimation using the published data. For reasonable loss measures, the trade-off is hyperbolic in character. Some encouraging simulation results are reported.