Precise Approximation of Convolutional NeuralNetworks for Homomorphically Encrypted Data

Many industries with convolutional neural network models offer privacy-preserving machine learning (PPML) classification service, that is, a service that performs classification of private data for clients while guaranteeing privacy. This work aims to study deep learning on the encrypted data using fully homomorphic encryption (FHE). To implement deep learning on FHE, ReLU and max-pooling functions should be approximated by some polynomials for homomorphic operations. Since the approximate polynomials studied up to now have large errors in ReLU and max-pooling functions, using these polynomials requires many epochs retraining for the classification of small datasets such as MNIST and CIFAR-10. In addition, large datasets such as ImageNet cannot be classified with high accuracy even with many epoch retraining using these polynomials. To overcome these disadvantages, we propose a precise polynomial approximation technique for ReLU and max-pooling functions. Since precise approximation requires a very high-degree polynomial, which may cause large numerical errors in FHE, we propose a method to approximate ReLU and max-pooling functions accurately using a composition of minimax approximate polynomials of small degrees. If we replace the ReLU and max-pooling functions with the proposed approximate polynomials, deep learning models such as ResNet and VGGNet, which have already been studied a lot, can still be used without further modification for PPML on FHE, and even pretrained parameters can be used without retraining. When we approximate ReLU function in the ResNet-152 using the composition of minimax approximate polynomials of degrees 15, 27, and 29, we succeed in classifying the plaintext ImageNet dataset for the first time with 77.52% accuracy, which is very close to the original model accuracy of 78.31%.