GPU Acceleration of Sparse Fully Homomorphic Encrypted DNNs

TL;DR

This paper presents a GPU-accelerated method for efficient sparse matrix multiplication in fully homomorphic encryption, significantly improving performance and complexity for deep neural network computations.

Contribution

It introduces a novel GPU-based sparse FHE matrix multiplication technique that outperforms CPU methods and reduces computational complexity.

Findings

GPU implementation outperforms CPU by up to 3.0×

Reduces time complexity from cubic to semi-linear

Demonstrates practical acceleration for FHE-based neural network inference

Abstract

Fully homomorphic encryption (FHE) has recently attracted significant attention as both a cryptographic primitive and a systems challenge. Given the latest advances in accelerated computing, FHE presents a promising opportunity for progress, with applications ranging from machine learning to information security. We target the most computationally intensive operation in deep neural networks from a hardware perspective, matrix multiplication (matmul), and adapt it for execution on AMD GPUs. We propose a new optimized method that improves the runtime and complexity of ciphertext matmul by using FIDESlib, a recent open-source FHE library designed specifically for GPUs. By exploiting sparsity in both operands, our sparse matmul implementation outperforms its CPU counterpart by up to $3.0\times$ and reduces the time complexity from cubic to semi-linear, demonstrating an improvement over existing FHE matmul implementations.