Personalized differential privacy Sample Clauses

Personalized differential privacy. (Chapter 5). This chapter focuses on personalized privacy setting, in which record owners in a dataset can have different expectations regarding the acceptable level of privacy for their data. One possibility for achieving PDP is to uses the minimal privacy budget among all records, but

Personalized differential privacy. We measure the count query accuracy by the relative frequency error defined as follows: for a query q, Aact(q) is the true answer to q on the original data. Anoisy(q) denotes the answer to q when using data generated from partitioning-based mecha- nisms or the sampling mechanism. Then the relative frequency error is defined as: RFE(q) = Aact(q) − Anoisy(q) where n is the total number of records in the original dataset. The reason we use relative frequency error is that the sampling mechanism generates a partial number of records from original datasets, and we need the query answers to be measured on the same base of record number. For the support vector machine, we use the area under the curve (AUC), and higher AUC value means better discrimination. For logistic regression, we convert Annual Income into a binary attribute: values higher than a predefined threshold are mapped to 1, and 0 otherwise. Accordingly, when we use a logistic model to exp(xT w) classify a tuple t, we predict the Annual Income of t to be 1 if 1+exp(xT w) > 0.5, where w is the model parameter, and x is a vector that contains the values of t on all attributes except Annual Income. We measure the accuracy of a logistic model by its misclassification rate, i.e., the fraction of tuples that are incorrectly classified.

Personalized differential privacy. Figure 5.7. US: (SVM) AUC for support vector machine classifica- tion under uniform privacy preferences, as privacy budget threshold are varied.