Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

High-Resolution Convolutional Neural Networks on Homomorphically Encrypted Data via Sharding Ciphertexts

Vivian Maloney, Richard F. Obrecht, Vikram Saraph, Prathibha Rama, Kate Tallaksen

TL;DR

This work demonstrates high-resolution DCNN inference on encrypted data by introducing sharding-based image and channel partitioning, enabling wide and deep networks to operate under the RNS-CKKS homomorphic encryption framework. It develops efficient convolution, pooling, and layer implementations that support large inputs and channels while maintaining accuracy, including three ResNet families that achieve 80.2% top-1 on ImageNet and 98.3% on CIFAR-10 with encrypted inference speeds 4.6–6.5× faster than prior methods. A key contribution is the training-time kurtosis regularization that constrains activation ranges, enabling low-degree polynomial approximations of GELU and reducing bootstrapping noise; this is complemented by Chebyshev-based polynomial approximations achieving depth-efficient evaluation. The result is the highest reported encrypted accuracies on CIFAR-10 and ImageNet-1k for PPML with FHE, alongside practical latency improvements that move encrypted inference closer to real-world privacy-preserving deployment. The work also outlines future directions toward encrypted object detection and transformer-based privacy-preserving analytics, and discusses potential hardware acceleration to further reduce latency.

Abstract

Recently, Deep Convolutional Neural Networks (DCNNs) including the ResNet-20 architecture have been privately evaluated on encrypted, low-resolution data with the Residue-Number-System Cheon-Kim-Kim-Song (RNS-CKKS) homomorphic encryption scheme. We extend methods for evaluating DCNNs on images with larger dimensions and many channels, beyond what can be stored in single ciphertexts. Additionally, we simplify and improve the efficiency of the recently introduced multiplexed image format, demonstrating that homomorphic evaluation can work with standard, row-major matrix packing and results in encrypted inference time speedups by $4.6-6.5\times$. We also show how existing DCNN models can be regularized during the training process to further improve efficiency and accuracy. These techniques are applied to homomorphically evaluate a DCNN with high accuracy on the high-resolution ImageNet dataset, achieving $80.2\%$ top-1 accuracy. We also achieve an accuracy of homomorphically evaluated CNNs on the CIFAR-10 dataset of $98.3\%$.

High-Resolution Convolutional Neural Networks on Homomorphically Encrypted Data via Sharding Ciphertexts

TL;DR

This work demonstrates high-resolution DCNN inference on encrypted data by introducing sharding-based image and channel partitioning, enabling wide and deep networks to operate under the RNS-CKKS homomorphic encryption framework. It develops efficient convolution, pooling, and layer implementations that support large inputs and channels while maintaining accuracy, including three ResNet families that achieve 80.2% top-1 on ImageNet and 98.3% on CIFAR-10 with encrypted inference speeds 4.6–6.5× faster than prior methods. A key contribution is the training-time kurtosis regularization that constrains activation ranges, enabling low-degree polynomial approximations of GELU and reducing bootstrapping noise; this is complemented by Chebyshev-based polynomial approximations achieving depth-efficient evaluation. The result is the highest reported encrypted accuracies on CIFAR-10 and ImageNet-1k for PPML with FHE, alongside practical latency improvements that move encrypted inference closer to real-world privacy-preserving deployment. The work also outlines future directions toward encrypted object detection and transformer-based privacy-preserving analytics, and discusses potential hardware acceleration to further reduce latency.

Abstract

Recently, Deep Convolutional Neural Networks (DCNNs) including the ResNet-20 architecture have been privately evaluated on encrypted, low-resolution data with the Residue-Number-System Cheon-Kim-Kim-Song (RNS-CKKS) homomorphic encryption scheme. We extend methods for evaluating DCNNs on images with larger dimensions and many channels, beyond what can be stored in single ciphertexts. Additionally, we simplify and improve the efficiency of the recently introduced multiplexed image format, demonstrating that homomorphic evaluation can work with standard, row-major matrix packing and results in encrypted inference time speedups by . We also show how existing DCNN models can be regularized during the training process to further improve efficiency and accuracy. These techniques are applied to homomorphically evaluate a DCNN with high accuracy on the high-resolution ImageNet dataset, achieving top-1 accuracy. We also achieve an accuracy of homomorphically evaluated CNNs on the CIFAR-10 dataset of .
Paper Structure (32 sections, 15 equations, 4 figures, 5 tables)

This paper contains 32 sections, 15 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Homomorphic encyption
  4. Threat Model
  5. Related Work
  6. Homomorphic Neural Network Operators
  7. Sharding and Encrypting an Image
  8. Duplicating and Permuting Channels
  9. Convolution
  10. With a Single Shard
  11. With Image Shards
  12. Single Shard with Duplication and Permutation
  13. With Channel Shards
  14. Average Pooling
  15. With Image Shards
  16. ...and 17 more sections

Figures (4)

  • Figure 1: Illustrations of image sharding and channel sharding.
  • Figure 2: (a) Partial convolution computation for a $4$-channel image convolved with a $1 \times 1$ kernel. (b) A single convolution computed by shifting the matrix. (c) Shifting rows from channel shards into adjacent ones.
  • Figure 3: Steps involved in a pooling operation. Duplication is not depicted.
  • Figure 4: The inputs to all GELU functions for the ResNet-20 architecture and the CIFAR-10 test set. The vertical dotted lines represent a proposed GELU bound.