Semantically Secure Classifiers for Privacy Preserving Data Mining

Abstract

Essential predictions are to be made by the parties distributed at multiple locations. However, in the process of building a model, perceptive data is not to be revealed. Maintaining the privacy of such data is a foremost concern. Earlier approaches developed for classification and prediction are proven not to be secure enough and the performance is affected. This chapter focuses on the secure construction of commonly used classifiers. The computations performed during model building are proved to be semantically secure. The homomorphism and probabilistic property of Paillier is used to perform secure product, mean, and variance calculations. The secure computations are performed without any intermediate data or the sensitive data at multiple sites being revealed. It is observed that the accuracy of the classifiers modeled is almost equivalent to the non-privacy preserving classifiers. Secure protocols require reduced computation time and communication cost. It is also proved that proposed privacy preserving classifiers perform significantly better than the base classifiers.