HEFactory: A symbolic execution compiler for privacy-preserving Deep Learning with Homomorphic Encryption

Homomorphic Encryption (HE) allows computing operations on encrypted data, and it is a potential solution to enable Deep Learning (DL) in privacy-enforcing scenarios (e.g., sending private data to cloud services). However, HE remains a complex technology with multiple challenges that prevent successful application by non-experts. In this work, we present HEFactory, a program compiler that effectively assists in building HE applications in Python for both general-purpose and Deep Learning applications, focusing on non-expert data scientists. HEFactory relies on a layered architecture that deals with challenges such as automatic parameter selection and specific data representation of HE applications. Our benchmarks show that HEFactory substantially lowers the programming complexity (i.e., a reduction of 80% in the number of lines of code) with negligible performance overhead over programs written by experts using native HE frameworks.(c) 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).