Precise Approximation of Convolutional Neural Networks for Homomorphically Encrypted Data

Homomorphic encryption (HE) is one of the representative solutions to privacy-preserving machine learning (PPML) classification enabling server classify private data clients while guaranteeing privacy. This work focuses on PPML using word-wise fully homomorphic (FHE). In order implement deep HE, ReLU and max-pooling functions should be approximated by polynomials for operations. Most previous studies focus HE-friendly networks, which approximate low-degree polynomials. However, this approximation cannot support deeper neural networks due large errors in general can only relatively small datasets. Thus, we propose a precise polynomial technique, composition minimax low degrees functions. If replace with proposed polynomials, standard models such as ResNet VGGNet still used without further modification FHE. Even pre-trained parameters retraining, makes method more practical. We ResNet-152 15, 27, 29. Then, succeed classifying plaintext ImageNet dataset 77.52% accuracy, very close original model accuracy 78.31%. Also, obtain an 87.90% encrypted CIFAR-10 ResNet-20 any additional training.