摘要
同态加密可以满足计算外包、数据共享、数据交易等应用对隐私保护计算的需要,但是同态加密的高计算开销限制了它在实际生产中的应用.本文从硬件卸载的角度解决同态加密的高性能计算问题,基于Intel QAT加速卡实现了一个半同态加密的高性能异步卸载框架QHCS. QHCS通过重构同态加密应用的软件栈来实现高效的异步卸载,并通过引入协程机制、批量加密技术等实现加密性能的最大化.本文同时给出了偏好不同性能指标(吞吐量、延迟)的两种卸载方案.进一步地,在由GPU及QAT组成的异构计算系统中,利用QHCS完整地实现了一个隐私保护的线性回归应用.实验结果表明,QHCS的吞吐量是目前软件实现的110倍,在百万量级的高维数据上实施隐私保护的线性回归计算只需十几分钟,可以较好地满足实际应用的需要.
Homomorphic encryption can preserve data privacy in applications such as outsourced computation,data sharing,and data trading,but the prohibited high computational overhead limits its application in real production.This paper tries to tackle the problem from the perspective of hardware offloading,and based on the Intel QuickAssist accelerator realize a high-performance asynchronous offloading framework called QHCS for a partially homomorphic encryption algorithm.QHCS achieves efficient asynchronous offloading by reconstructing the software stack of homomorphic encryption applications,and then maximizes encryption performance by introducing fibre mechanisms,batch encryption technologies and so on.This paper also presents two offloading schemes that prefer different performance metrics(throughput and latency).Furthermore,a privacy-preserving linear regression application is implemented with QHCS in a heterogeneous computing system comprising GPU and QAT.The experimental results show that QHCS achieves an encryption throughput 110 times as high as that of the existing software implementation,and only takes a dozen minutes to finish the privacypreserving linear regression on one million pieces of high-dimensional data.The latter experiment demonstrates that QHCS can meet the requirement of practical applications.
作者
周慧凯
华蓓
ZHOU Hui-kai;HUA Bei(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230027,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第3期595-600,共6页
Journal of Chinese Computer Systems
关键词
隐私保护计算
同态加密
硬件卸载
QAT加速器
privacy-preserving computing
homomorphic encryption
hardware offloading
QAT accelerator
