GPU Acceleration of High-Precision Homomorphic Computation Utilizing Redundant Representation

Fully homomorphic encryption (FHE) can perform computations on encrypted data, allowing us to analyze sensitive data without losing its security. The main issue for FHE is its lower performance, especially for high-precision computations, compared to calculations on plaintext data. Making FHE viable for practical use requires both algorithmic improvements and hardware acceleration. Recently, Klemsa and Onen (CODASPY'22) presented fast homomorphic algorithms for high-precision integers, including addition, multiplication and some fundamental functions, by utilizing a technique called redundant representation. Their algorithms were applied on TFHE, which was proposed by Chillotti et al. (Asiacrypt'16). In this paper, we further accelerate this method by extending their algorithms to multithreaded environments. The experimental results show that our approach performs 128-bit addition in 0.41 seconds, 32-bit multiplication in 4.3 seconds, and 128-bit Max and ReLU functions in 1.4 seconds using a Tesla V100S server.