Risk-Efficient Bayesian Data Synthesis for Privacy Protection

Abstract Statistical agencies utilize models to synthesize respondent-level data for release the public privacy protection. In this study, we efficiently induce protection into any Bayesian synthesis model by employing a pseudo-likelihood that exponentiates each likelihood contribution an observation record-indexed weight ∈[0,1], defined be inversely proportional identification risk record. We start with marginal probability of record, which is composed as identity record may disclosed. Our application Consumer Expenditure Surveys (CE) U.S. Bureau Labor Statistics demonstrates marginally risk-weighted synthesizer provides overall improved However, risks actually increase some moderate-risk records after pseudo-posterior estimation owing increased isolation weighting, phenomenon label “whack-a-mole.” proceed construct from collection pairwise probabilities other records, where measures joint reidentification pair mitigates whack-a-mole issue and produces more efficient set synthetic lower higher utility CE data.