Efficient homomorphic encryption framework for privacy-preserving regression

Abstract

Homomorphic encryption (HE) has recently attracted considerable attention as a key solution for privacy-preserving machine learning because HE can apply to various areas that require to delegate outsourcing computations of user's data. Nevertheless, its computational inefficiency still hinders its wider application. In this study, we propose an alternative to bridge the gap between the privacy and efficiency of HE by encrypting only a small amount of private information. We first derive an exact solution to HE-friendly ridge regression with multiple private variables, while linearly reducing the computational complexity of this algorithm over the number of variables. The proposed method has the advantage that it can be implemented using any HE scheme. Moreover, we propose an adversarial perturbation method that can prevent potential attacks on private variables, which have rarely been explored in HE-based machine learning studies. An extensive experiment on real-world benchmarking datasets supports the effectiveness of our method.