A Partitioned Neural Network Architecture for Efficient Inference with Fully Homomorphic Encryption

Abstract

: Fully homomorphic encryption (FHE) enables computations on ciphertexts without decryption. This property is expected to be utilized in AI with sensitive data. Although encryption improves the security of neural network inference, it incurs a significant computational overhead because all processes are executed under encryption. However, in many practical scenarios, not necessarily all the input features should be encrypted. Some features must be confidential whereas others can be disclosed to the model operator. Based on this observation, we propose a novel neural network architecture which partitions the input features into two types according to their secrecy. Our architecture decomposes a neural network into three modules to handle these two feature types efficiently. Public input features are processed without encryption whereas private input features are computed under encryption. We theoretically analyze the computational cost of our model and formulate the reduction rate in terms of the parameters. We also experimentally examine our model’s accuracy by comparing it to that of a standard model and demonstrate that our model reduces the computational costs by 50% for a certain parameter set while the accuracy degradation is limited.