Differentially private nearest neighbor classification

Instance-based learning, and the k-nearest neighbors algorithm (k-NN) in particular, provide simple yet effective classification algorithms for data mining. Classifiers are often executed on sensitive information such as medical or personal data. Differential privacy has recently emerged as the accepted standard for privacy protection in sensitive data. However, straightforward applications of differential privacy to k-NN classification yield rather inaccurate results. Motivated by this, we develop algorithms to increase the accuracy of private instance-based classification. We first describe the radius neighbors classifier (r-N) and show that its accuracy under differential privacy can be greatly improved by a non-trivial sensitivity analysis. Then, for k-NN classification, we build algorithms that convert k-NN classifiers to r-N classifiers. We experimentally evaluate the accuracy of both classifiers using various datasets. Experiments show that our proposed classifiers significantly outperform baseline private classifiers (i.e., straightforward applications of differential privacy) and executing the classifiers on a dataset published using differential privacy. In addition, the accuracy of our proposed k-NN classifiers are at least comparable to, and in many cases better than, the other differentially private machine learning techniques.