Uma avaliação de abordagens LDP aplicadas a conjuntos de dados longitudinais

Abstract

Local differential privacy (LDP) was developed as a more strict version of differential privacy (DP), the state-of-the-art model of anonymity guarantee for databases. Due to the requirement of data anonymity before it is sent to the server, a challenge in guaranteeing LDP is the risk of excessive addition of noise to the data, which can be especially hard to avoid when applying LDP to longitudinal data, that requires successive queries over time, with each one having to guarantee LDP. In this paper, we aim to evaluate the performance of LDP protocols adapted for the protection of longitudinal data, when applied to the task of finding the k most frequent items, and their frequencies, among longitudinal data sets. To this end, we evaluate the performance of a wide range of LDP mechanisms when used in conjunction with the state-of-the-art SVIM approach, in the processing of four different datasets. After exhaustive experimentation, we compared the results found and indicated the most promising ones.