面向深度神经网络的安全计算协议设计方法

doi:10.11959/j.issn.2096-109x.2020050

网络与信息安全学报 ›› 2020, Vol. 6 ›› Issue (4): 130-139.doi: 10.11959/j.issn.2096-109x.2020050

面向深度神经网络的安全计算协议设计方法

毕仁万¹,陈前昕¹,熊金波¹(),刘西蒙²

¹ 福建师范大学数学与信息学院，福建福州 350117
² 福州大学数学与计算机科学学院，福建福州 350108

修回日期:2020-07-03 出版日期:2020-08-15 发布日期:2020-08-13
作者简介:毕仁万（1996- ），男，湖南常德人，福建师范大学硕士生，主要研究方向为安全深度学习、安全多方计算|陈前昕（1996- ），男，福建泉州人，福建师范大学硕士生，主要研究方向为安全深度学习、隐私保护技术|熊金波（1981- ），男，湖南益阳人，博士，福建师范大学教授，主要研究方向为安全深度学习、移动群智感知、隐私保护技术|刘西蒙（1988- ），男，陕西西安人，博士，福州大学教授，主要研究方向为云安全、应用密码学、大数据安全
基金资助:
国家自然科学基金(61872088);国家自然科学基金(U1804263);国家自然科学基金(61702105);国家自然科学基金(61872090);福建省自然科学基金资助项目(2019J01276);贵州省公共大数据重点实验室开放课题(2019BDKFJJ004)

Design method of secure computing protocol for deep neural network

Renwan BI¹,Qianxin CHEN¹,Jinbo XIONG¹(),Ximeng LIU²

¹ College of Mathematics and Informatics,Fujian Normal University,Fuzhou 350117,China
² College of Mathematics and computer science,Fuzhou University,Fuzhou 350108,China

Revised:2020-07-03 Online:2020-08-15 Published:2020-08-13
Supported by:
The National Natural Science Foundation of China(61872088);The National Natural Science Foundation of China(U1804263);The National Natural Science Foundation of China(61702105);The National Natural Science Foundation of China(61872090);The Natural Science Foundation of Fujian Province,China(2019J01276);The Guizhou Provincial Key Laboratory of Public Big Data Research Fund(2019BDKFJJ004)

摘要/Abstract

摘要：

针对深度神经网络模型计算过程中存在的信息泄露问题，结合加性秘密共享方案，在两台非共谋的边缘服务器间设计安全高效的交互计算协议。考虑到非线性函数不能直接拆分，首先提出一组基本转换协议，实现加性副本和乘性副本的安全转换，经过少量调用，可以安全计算幂函数、比较、指数、对数、除法等底层函数。由于数据传递和计算特点，协议可以扩展至数组计算。理论分析证明了协议的正确性、高效性和安全性，实验结果表明，协议具有较小的误差，其计算和通信开销均优于现有设计方案。

关键词: 深度神经网络, 加性秘密共享, 安全计算协议, 加法-乘法转换, 数组单元

Abstract:

Aiming at the information leakage problem in the process of deep neural network model calculation,a series of secure and efficient interactive computing protocols were designed between two non-collusive edge servers in combination with the additive secret sharing scheme.Since the nonlinear function cannot be split directly,a set of basic conversion protocols were proposed to realize the secure conversion of additive and multiplicative shares.After a few invokes,the power,comparison,exponential,logarithm,division and other low-level functions can be calculated securely.Due to the characteristics of data transfer and computation,the proposed protocols can be extended to array computation.Theoretical analysis ensures the correctness,efficiency and security of these protocols.The experimental results show that the error of these protocols is negligible,and the computational costs and communication overhead are better than the existing schemes.

Key words: deep neural network, additive secret sharing, secure computing protocol, add-multiply transformation, array unit

中图分类号:

TP393

毕仁万,陈前昕,熊金波,刘西蒙. 面向深度神经网络的安全计算协议设计方法[J]. 网络与信息安全学报, 2020, 6(4): 130-139.

Renwan BI,Qianxin CHEN,Jinbo XIONG,Ximeng LIU. Design method of secure computing protocol for deep neural network[J]. Chinese Journal of Network and Information Security, 2020, 6(4): 130-139.

图/表 6

图1

表1

表2

表3

表4

表5

参考文献 18

[1]	ZHOU Y , TUZEL O . Voxelnet:end-to-end learning for point cloud based 3d object detection[C]// The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2018: 4490-4499.
[2]	REN S Q , HE K , GIRSHICK R ,et al. Faster R-CNN:towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015,39(6): 91-99.
[3]	YANN L , BERNHARD B , JOHN S D ,et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Computation, 1989,1(4): 541-551.
[4]	XIONG J B , BI R W , SZHAO M F ,et al. Edge-assisted privacy-preserving raw data sharing framework for connected autonomous vehicles[J]. IEEE Wireless Communications, 2020,27(3): 24-30.
[5]	HESAMIFARD E , TAKABI H , GHASEMI M . CryptoDL:deep neural networks over encrypted data[J]. arXiv preprint arXiv:1711.05189, 2017
[6]	MOHASSEL P , ZHANG Y P . SecureML:a system for scalable privacy-preserving machine learning[C]// IEEE Symposium on Security and Privacy (S＆P). 2017: 19-38.
[7]	YAO A C , . Protocols for secure computations[C]// The 23rd Annual Symposium on Foundations of Computer Science (SFCS). 1982: 160-164.
[8]	唐春明, 魏伟明 . 基于安全两方计算的具有隐私性的回归算法[J]. 信息网络安全, 2018(10): 10-16.
	TANG C M , WEI W M . Regression algorithm with based on secure two-party computation[J]. Information Network Security, 2018(10): 10-16.
[9]	RABIN M O . How to exchange secrets with oblivious transfer[J]. IACR Cryptology ePrint Archive, 2005:187.
[10]	刘新, 李顺东, 陈振华 . 基本初等函数的保密云计算服务协议[J]. 计算机科学, 2015,42(10): 159-163.
	LIU X , LI S D , CHEN Z H . Secure cloud computing service protocols of elementary functions[J]. Computer Science, 2015,42(10): 159-163.
[11]	HUANG K , LIU X M , FU S J ,et al. A lightweight privacy-preserving CNN feature extraction framework for mobile sensing[J]. IEEE Transactions on Dependable and Secure Computing, 2019:1.
[12]	MA Z , LIU Y , LIU X M ,et al. Privacy-preserving outsourced speech recognition for smart IoT devices[J]. IEEE Internet of Things Journal, 2019,6(5): 8406-8420.
[13]	LIU Y , MA Z , LIU X M ,et al. Privacy-preserving object detection for medical images with Faster R-CNN[J]. IEEE Transactions on Information Forensics and Security, 2019:1.
[14]	VOLDER J E . The cordic trigonometric computing technique[J]. IRE Transactions on Electronic Computers, 1959,8(3): 330-334.
[15]	ZHU Y W , HUANG L S , YANG W ,et al. Three new approaches to privacy-preserving add to multiply protocol and its application[C]// The 2nd International Workshop on Knowledge Discovery and Data Mining. 2009: 554-558.
[16]	李禾, 王述洋 . 关于除法的安全双方计算协议[J]. 计算机工程与应用, 2010,46(6): 86-88.
	LI H , WANG S Y . Several secure two-party protocols of division[J]. Computer Engineering and Applications, 2010,46(6): 86-88.
[17]	BOGDANOV D , LAUR S , WILLEMSON J . Sharemind:a framework for fast privacy-preserving computations[C]// European Symposium on Research in Computer Security. 2008: 192-206.
[18]	BOGDANOV D , NIITSOO M , TOFT T ,et al. High-performance secure multi-party computation for data mining applications[J]. International Journal of Information Security, 2012,11(6): 403-418.

相关工作	SPow	SRec	SMul	SComp	SExp	SLog	SDiv
文献[11]	N.A.	N.A.	O (n)	O(nl)	N.A.	N.A.	N.A.
文献[12]	O(nm ²)	O (nm)	N.A.	N.A.	O (nm)	O (nm)	N.A.
文献[16]	N.A.	N.A.	N.A.	N.A.	N.A.	N.A.	O (n)
本文	O (n)	O (n)	O (n)	O (n)	O (n)	O (n)	O (n)

相关工作	SPow	SRec	SMul	SComp	SExp	SLog	SDiv
文献[11]	N.A.	N.A.	1	2l+4	N.A.	N.A.	N.A.
文献[12]	2m ²+m	2m	N.A.	N.A.	m+ 1	3m+2	N.A.
文献[16]	N.A.	N.A.	N.A.	N.A.	N.A.	N.A.	5
本文	3	3	2	3	1	2	5

	输入范围	1 ¹0	1 ²0	1 ³0	1 ⁴0
STMA	AE	10^-7	10^-6	10^-5	10^-4
	RE	10^-6	10^-7	10^-7	10^-8
STAM	AE	10^-5	10^-5	10^-5	10^-5
	RE	10^-4	10^-5	10^-7	10^-8
SPow	AE	10^-5	10^-6	10^-6	10^-7
	RE	10^-4	10^-6	10^-7	10^-8
SMul	AE	10^-7	10^-6	10^-5	10^-4
	RE	10^-7	10^-8	10^-10	10^-9
SLog	AE	10^-3	10^-4	10^-4	10^-5
	RE	10^-3	10^-4	10^-5	10^-6

输入数量n							运行时间/ms
输入数量n	STMA	STAM	SPow	SRec	SMul^[11]	SMul	SComp^[11]	SComp	SExp	SLog	SDiv ^[16]	SDiv
1 ⁴0	≈0	0.10	0.10	0.10	0.20	0.10	7.88	0.20	0.10	0.10	0.30	0.10
1 ⁵0	0.60	0.70	1.70	1.40	0.99	1.70	39.20	1.20	0.80	0.80	3.39	2.09
1 ⁶0	8.28	8.08	19.64	16.25	11.87	20.85	481.71	13.97	10.18	8.87	43.08	25.73
1 ⁷0	87.17	91.16	206.95	166.85	110.61	214.63	4770.44	147.51	106.02	92.15	417.18	252.13

参与方							通信开销/kB
参与方	STMA	STAM	SPow	SRec	SMul^[11]	SMul	SComp^[11]	SComp	SExp	SLog	SDiv ^[16]	SDiv
S₁	0.76	0.76	1.53	1.53	1.53	1.53	24.41	1.53	0.76	0.76	3.81	2.29
S₂	0.76	1.53	2.29	2.29	1.53	1.53	24.41	2.29	0.76	1.53	3.81	3.81
T	1.53	1.53	3.05	3.05	2.29	3.05	1.53	1.53	1.53	1.53	6.10	4.58

面向深度神经网络的安全计算协议设计方法

Design method of secure computing protocol for deep neural network

在线阅读

pdf下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 6

参考文献 18

相关文章 3

Metrics

推荐阅读 0

[1]	陈浩, 王锋, 张卫明, 俞能海. 载体独立的深度光照水印算法[J]. 网络与信息安全学报, 2022, 8(4): 110-118.
[2]	超凡,杨智,杜学绘,孙彦. 基于深度神经网络的Android恶意软件检测方法[J]. 网络与信息安全学报, 2020, 6(5): 67-79.
[3]	严飞,张铭伦,张立强. 基于边界值不变量的对抗样本检测方法[J]. 网络与信息安全学报, 2020, 6(1): 38-45.