Uma abordagem de privacidade diferencial para dados correlacionados utilizando técnicas de agrupamento

Abstract

Differential Privacy is a mathematical model designed to hinder the process of distinguishing individuals’ records on statistical databases, while maximizing data utility. Although Differential Privacy has been widely used for protecting the privacy of individual users’ data, it was not designed to provide its guarantees for correlated data, since it considers, in essence, independence of records in the database. Existing techniques using Differential Privacy on correlated data attempt to use dependence parameters or correlation coefficients (such as Pearson or Spearman’s Rank) to measure the correlation among records in a dataset. However, they tend to introduce an amount of noise higher than the necessary in the query answer, decreasing the data utility. Different from the existing works, we propose an approach that clusters similar records, which are more likely to be correlated, based on Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Gaussian Mixture Model (GMM). Our approach also employs a correlated Laplace mechanism to compute the privatized answers, satisfying the privacy guarantees of Differential Privacy. The experimental evaluation exhibits the benefits of our clustering strategy in terms of effectiveness and efficiency, considering data utility and privacy.