Research And Implementation Of Heterogeneous Computing Architectures For Privacy-Preserving Machine Learning Inference Acceleration

With the development of computer technology,especially the rise of artificial intelligence,machine learning has played a huge role in the fields of finance,e-commerce,and healthcare.However,once the raw data of machine learning is leaked,it will bring many adverse effects.Existing research shows that attackers can restore the original data from machine learning models through attacks.Therefore,how to perform machine learning while protecting data privacy becomes a critical issue.And protecting the privacy of machine learning by cryptographic techniques usually has a large computational and communication overhead.In view of this background,this topic develops a heterogeneous computing platform for privacy-preserving machine learning with accelerated inference phase based on Yao’s protocol to address the problem of low computational efficiency of privacy-preserving machine learning.The main work is as follows:(1)An overlay architecture based on CPU-FPGA heterogeneous computing is implemented,including the garbled circuit generation side and the proxy computing side.The architecture accelerates the garbled circuit generation phase and the computation phase by taking advantage of the CPU’s expertise in processing control flow and the FPGA’s parallel processing,which improves the efficiency of generating and computing garbled circuits and reduces the computational pressure.(2)A lightweight proxy oblivious transfer protocol is proposed.Compared with the oblivious transfer protocol implemented by asymmetric cryptographic algorithms,in the lightweight proxy oblivious transfer protocol,the generator of the garbled circuit and the proxy calculator only need to perform symmetric operations,and the proxy calculator can obtain the random number held by the generator corresponding to the user input.This lightweight proxy oblivious transfer protocol reduces the computational pressure on the user and the server during the oblivious transfer phase.(3)Experiments are designed to verify the effectiveness of the above scheme.In a LAN environment,the scheme is 2.2 times more efficient in generating and computing garbled circuits compared to the software implementation of the Yao’s protocol.Compared to the oblivious transfer protocol,the lightweight proxy oblivious transfer protocol is 398 times more efficient.Moreover,a heterogeneous computing platform for privacy-preserving machine learning inference acceleration is designed and developed.After deployment,the platform is capable of accelerating privacy-preserving machine learning inference using heterogeneous computing.