Paper 2024/1948

ARK: Adaptive Rotation Key Management for Fully Homomorphic Encryption Targeting Memory Efficient Deep Learning Inference

Abstract

Advancements in deep learning (DL) not only revolutionized many aspects in our lives, but also introduced privacy concerns, because it processed vast amounts of information that was closely related to our daily life. Fully Homomorphic Encryption (FHE) is one of the promising solutions to this privacy issue, as it allows computations to be carried out directly on the encrypted data. However, FHE requires high computational cost, which is a huge barrier to its widespread adoption. Many prior works proposed techniques to enhance the speed performance of FHE in the past decade, but they often impose significant memory requirements, which may be up to hundreds of gigabytes. Recently, focus has shifted from purely improving speed performance to managing FHE’s memory consumption as a critical challenge. Rovida and Leporati introduced a technique to minimize rotation key memory by retaining only essential keys, yet this technique is limited to cases with symmetric numerical patterns (e.g., -2 -1 0 1 2), constraining its broader utility. In this paper, a new technique, Adaptive Rotation Key (ARK), is proposed that minimizes rotation key memory consumption by exhaustively analyzing numerical patterns to produce a minimal subset of shared rotation keys. ARK also provides a dual-configuration option, enabling users to prioritize memory efficiency or computational speed. In memory-prioritized mode, ARK reduces rotation key memory consumption by 41.17% with a 12.57% increase in execution time. For speed-prioritized mode, it achieves a 24.62% rotation key memory reduction with only a 0.21% impact on execution time. This flexibility positions ARK as an effective solution for optimizing FHE across varied use cases, marking a significant advancement in optimization strategies for FHE-based privacy-preserving systems.