With increasing awareness of data security and privacy in cloud-based services, homomorphic encryption (HE) has emerged as a promising solution for securely outsourcing computation. HE enables direct computation on encrypted data (ciphertext) without ...
With increasing awareness of data security and privacy in cloud-based services, homomorphic encryption (HE) has emerged as a promising solution for securely outsourcing computation. HE enables direct computation on encrypted data (ciphertext) without requiring decryption, thereby preventing potential privacy breaches during cloud processing.
However, since HE relies on noisy encryption schemes in which noise accumulates with each operation, early-stage HE schemes could support only a limited number of computations before the noise corrupted the data. To overcome this limitation, fully homomorphic encryption (FHE) introduces a bootstrapping operation that refreshes the accumulated noise without decrypting the data, thereby enabling an unlimited number of computations on encrypted data. Bootstrapping is an essential operation for applying HE to real-world workloads. Unfortunately, bootstrapping incurs substantial computational and memory bandwidth overhead, making HE less practical in many scenarios.
In this dissertation, we present hardware solutions to make HE more practical by focusing on accelerating the bootstrapping process. As a starting point, we accelerate the Number Theoretic Transform (NTT)—a core primitive responsible for the majority of computational overhead in HE operations—on a popular computing platform: the GPU.
With
Through in-depth analysis, we identify that NTT computations on GPUs suffer significantly from memory bandwidth bottlenecks, particularly under large parameter settings required to support bootstrapping.
To address this, we propose a data reuse strategy that leverages the hierarchical memory system of modern GPUs to substantially reduce off-chip memory access.
Furthermore, we introduce efficient on-the-fly generation methods for the constants used in NTT, yielding additional reductions in memory access and improving overall performance.
However, in broad scope, computing bootstrapping on GPU still suffers from
high memory bandwidth bottleneck due to memory expansion caused by HE encryption.
To overcome this challenge, we propose BTS—a Bootstrappable, Technology-driven, Secure accelerator architecture for FHE. We identify key challenges in supporting bootstrapping within hardware accelerators and analyze the associated off-chip memory bandwidth and computational requirements. In particular, considering the limitations of modern memory technologies, we determine parameter sets for homomorphic encryption that are well-suited for efficient FHE acceleration.
Based on the insights from our analysis, we design BTS to exploit the inherent parallelism of HE operations by deploying a large grid of processing elements. We detail the architectural and microarchitectural design of BTS, including a network-on-chip (NoC) optimized for the deterministic communication patterns characteristic of FHE workloads. BTS achieves 5,556× and 1,306× speedups over CPU implementations on ResNet-20 and logistic regression, respectively, while occupying 373.6mm² of chip area and consuming up to 163.2W of power.