$\textsf{LoPub}$ : High-Dimensional Crowdsourced Data Publication With Local Differential Privacy

摘要: High-dimensional crowdsourced data collected from numerous users produces rich knowledge about our society; however, it also brings unprecedented privacy threats to the participants. Local differential (LDP), a variant of privacy, is recently proposed as state-of-the-art notion. Unfortunately, achieving LDP on high-dimensional publication raises great challenges in terms both computational efficiency and utility. To this end, based expectation maximization (EM) algorithm Lasso regression, we first propose efficient multi-dimensional joint distribution estimation algorithms with LDP. Then, develop local differentially private ( LoPub ) by taking advantage techniques. In particular, correlations among multiple attributes are identified reduce dimensionality data, thus speeding up learning process high Extensive experiments real-world datasets demonstrate that multivariate scheme significantly outperforms existing schemes communication overhead speed. Moreover, can keep, average, 80% 60% accuracy over released support vector machine random forest classification, respectively.