MPLDP: Multi-Level Personalized Local Differential Privacy Method

Users have different sensitivities to different attributes for the same data set. Disregarding this can result in inadequate data confidentiality or reduced data availability. To address this, this paper proposes a multi-level personalized local differential privacy mechanism optimization method. In high-dimensional heterogeneous data scenario, this paper first adopts the optimal privacy budget allocation scheme to allocate the privacy budget of different attributes, and then categorizes the privacy levels into high, medium, and low. Users can freely select the privacy level for each attribute or choose the same level for all attributes. For data analysts, reorganizing data with different privacy levels to achieve histogram estimation is a challenging task. The paper introduces a histogram optimization estimation method based on two evaluation criteria. It proposes a combinatorial optimization method, OC, which minimizes mean square error, and a combinatorial optimization method, OP, based on perturbation theory, which minimizes maximum error. The paper comprehensively studies the balance between data availability and privacy protection based on these two rules.