Private machine learning classification based on fully homomorphic encryption

Machine learning classification is an useful tool for trend prediction by analyzing big data. As supporting homomorphic operations over encrypted data without decryption, fully homomorphic encryption (FHE) contributes to machine learning classification without leaking user privacy, especially in the outsouring scenario. In this paper, we propose an improved FHE scheme based on HElib, which is a FHE library implemented based on Brakerski's FHE scheme. Our improvement focuses on two aspects. On the one hand, we first use the relinearization technique to reduce the ciphertext size, and then the modulus switching technique is used to reduce the modulus and decryption noise. On the other hand, we need no relinearization and modulus switching if there is additive homomorphic or no homomorphic operation in the multiplicative ciphertext's next homomorphic operation. Homomorphic comparison protocol, private hyperplane decision-based classification and private Naïve Bayes classification are implemented by additive homomorphic and multiplicative homomorphic first. In our homomorphic comparison protocol, the number of interactions is reduced from 3 to 1. We choose the proposed FHE scheme to implement private decision tree classification. Simulation results show that the efficiency of our FHE scheme and implementation of private decision tree classification are more efficient than other two schemes.