Differential Privacy for Classifier Evaluation

Abstract

Differential privacy provides powerful guarantees that individuals incur minimal additional risk by including their personal data in a database. Most work in differential privacy has focused on differentially private algorithms that produce models, counts, and histograms. Nevertheless, even with a classification model produced by a differentially private algorithm, directly reporting the classifier's performance on a database has the potential for disclosure. Thus, differentially private computation of evaluation metrics for machine learning is an important research area. We find effective mechanisms for area under the receiver-operating characteristic (ROC) curve and average precision.