Paper 2023/632

High-Throughput Deep Convolutional Neural Networks on Fully Homomorphic Encryption Using Channel-By-Channel Packing

Abstract

Secure Machine Learning as a Service is a viable solution where clients seek secure delegation of the ML computation while protecting their sensitive data. We propose an efficient method to securely evaluate deep standard convolutional neural networks based on CKKS fully homomorphic encryption, in the manner of batch inference. In this paper, we introduce a packing method called Channel-by-Channel Packing that maximizes the slot compactness and single-instruction-multipledata capabilities in ciphertexts. Along with further optimizations such as lazy rescaling, lazy Baby-Step Giant-Step, and ciphertext level management, we could significantly reduce the computational cost of standard ResNet inference. Simulation results show that our work has improvements in amortized time by 5.04× (from 79.46s to 15.76s) and 5.20×(from 455.56s to 87.60s) for ResNet-20 and ResNet-110, compared to the previous best results, resp. We also got a dramatic reduction in memory usage for rotation keys from several hundred GBs to 6.91GB, which is about 38× smaller than the previous result.