On Secure Neural Network Inferences Using Fully Homomorphic Encryption

Abstract

Fully homomorphic encryption (FHE) is a form of encryption that allows us to perform arbitrary computation on encrypted data. We can use FHE to encrypt inputs and perform a neural network inference without revealing the inputs. The encrypted inference results in a higher computation cost than on plaintexts, which causes a large computational overhead when evaluating a complex neural network. In this paper, we assess secure inferences based on FHE over various network architectures and hyper parameters, and investigate a trade-off between inference accuracy and computation time.