HDNN: Homomorphic Decryption Neural Network

Abstract

This study proposes a homomorphic decryption layer to construct a homomorphic decryption neural network (HDNN) for improving security and reducing computational time. The ciphertexts of inputs can be adopted as the values of neurons in the input layer of HDNN for estimating the plaintexts of outputs. The results showed the proposed method can obtain high performance and save computational resources.