Privacy-Preserving Deep Learning With Homomorphic Encryption: An Introduction

Abstract

Privacy-preserving deep learning with homomorphic encryption (HE) is a novel and promising research area aimed at designing deep learning solutions that operate while guaranteeing the privacy of user data. Designing privacypreserving deep learning solutions requires one to completely rethink and redesign deep learning models and algorithms to match the severe technological and algorithmic constraints of HE. This paper provides an introduction to this complex research area as well as a methodology for designing privacy-preserving convolutional neural networks (CNNs). This methodology was applied to the design of a privacy-preserving version of the well-known LeNet-1 CNN, which was successfully operated on two benchmark datasets for image classification. Furthermore, this paper details and comments on the research challenges and software resources available for privacy-preserving deep learning with HE.

EDLaaS:Fully Homomorphic Encryption over Neural Network Graphs for Vision and Private Strawberry Yield Forecasting

“…In particular, in future we intend to continue to work applying FHE to existing networks, such as transformers, which are SotA in sequence tasks. However, much work remains in mimicking certain functions of transformers in an FHE-compatible manner [ 12 ]. We also believe while we could compute privately, we can significantly improve the performance of the predictions themselves with better performing network architectures such as transformers.…”

“…Here, FHE graph parameterisation means deriving the FHE parameters from a graph, such as the computational depth and thus the parameters like the modulus size. There have been a few works that define FHE graph parameterisation, the most notable and similar of which is Microsoft Encrypted Vector Arithmetic (MS-EVA) [ 11 , 12 ]. MS-EVA uses Directed Acyclic Graphs (DAGs) to represent simple operations applied to some input constant.…”

“…Our contributions towards FHE deep learning, given the current state of the field and related works ( Section 2.2 ) are: Propose a new block-level automatic cyphertext parameterisation algorithm, which we call autoFHE. We also seek to showcase autoFHE in both regression and classification networks, which still appears to be a misunderstood and ongoing problem [ 12 ]. Provide and showcase open-source encrypted deep learning with a reproducible step-by-step example on an open dataset, in this case Fashion-MNIST, achieved through a dockerised Jupyter-lab container, such that others can readily and easily explore FHE with deep learning and verify our results.…”

“…Propose a new block-level automatic cyphertext parameterisation algorithm, which we call autoFHE. We also seek to showcase autoFHE in both regression and classification networks, which still appears to be a misunderstood and ongoing problem [ 12 ].…”

Privacy-Preserving Federated Learning Based On Partial Low-Quality Data

“…In order to address privacy leakage concerns in Federated learning, researchers have proposed various solutions leveraging privacy protection technologies such as differential privacy [13], homomorphic encryption [14], and secure multi-party computation [15]. Haokun Fang et al [16] proposed a Federated learning scheme based on the Paillier additive homomorphic algorithm.…”