Privacy-Preserving Machine Learning With Fully Homomorphic Encryption for Deep Neural Network

Fully homomorphic encryption (FHE) is a prospective tool for privacy-preserving machine learning (PPML). Several PPML models have been proposed based on various FHE schemes and approaches. Although are suitable as tools implementing models, previous FHE, such CryptoNet, SEALion, CryptoDL, limited to simple nonstandard types of models; they not proven be efficient accurate with more practical advanced datasets. Previous replaced non-arithmetic activation functions arithmetic instead adopting approximation methods did use bootstrapping, which enables continuous evaluations. Thus, could neither standard nor employ large numbers layers. In this work, we first implement the ResNet-20 model RNS-CKKS bootstrapping verify implemented CIFAR-10 dataset plaintext parameters. Instead replacing functions, state-of-the-art evaluate these ReLU Softmax, sufficient precision. Further, time, technique scheme in model, us an arbitrary deep encrypted data. We numerically that shows 98.43% identical results original non-encrypted The classification accuracy 92.43%±2.65%, quite close CNN (91.89%). It takes approximately 3 h inference dual Intel Xeon Platinum 8280 CPU (112 cores) 172 GB memory. believe opens possibility applying model.