基于CNN-MGU的侧信道攻击研究

doi:10.3969/j.issn.1671-1122.2022.08.007

摘要/Abstract

摘要：

基于深度学习的侧信道攻击对密码算法的安全性具有严重威胁,是学术界研究的热点之一。目前神经网络模型存在准确率低、鲁棒性差、收敛速度慢等问题,针对这些问题,文章结合卷积神经网络（Convolutional Neural Network,CNN）和最小门控单元（Minimal Gated Unit,MGU）,提出基于CNN-MGU的神经网络模型。该模型首先通过CNN层提取轨迹上的局部关键信息,然后利用MGU层充分学习局部关键信息在时间上的相互依赖关系恢复密钥,最后在完全同步与非同步的轨迹上对模型的性能进行验证。实验结果表明,与基于CNN、长短期记忆（Long Short-Term Memory,LSTM）网络的攻击方法相比,基于CNN-MGU模型的训练准确率分别提高了约5.6%、3.4%。当数据集中加入的抖动量从0增大至50、100时,基于CNN-MGU的神经网络模型的准确率仍达90%,鲁棒性强、收敛速度快。

关键词: 侧信道攻击, 深度学习, CNN, MGU

Abstract:

Side-channel attacks based on deep learning pose a serious threat to the security of cryptographic algorithms and become a research hotspot. At present, some network models have problems such as low accuracy, poor robustness, and slow convergence. Aimed at these problems, this paper proposed a neural network model based on CNN-MGU by combining the advantages of convolutional neural network (CNN) and minimal gated unit (MGU). Firstly, the key information was effectively extracted by the CNN layer. Secondly, the timing dependency was fully learned by the MGU layer, and the key was recovered by dividing and conquering. Thirdly, the performance of the model was verified on fully synchronization and asynchronous traces. The experimental results show that compared with attack based on CNN and long short-term memory(LSTM) network, the accuracy of method based on CNN-MGU is improved by about 5.6% and 3.4% respectively. When the amount of jitter added in the data set increases from 0 to 50 and 100 respectively, the accuracy of the CNN-MGU is still 90% accuracy, which has strong robustness and fast convergence speed.

Key words: side channel attack, deep learning, CNN, MGU

中图分类号:

TP309

高博, 陈琳, 严迎建. 基于CNN-MGU的侧信道攻击研究[J]. 信息网络安全, 2022, 22(8): 55-63.

GAO Bo, CHEN Lin, YAN Yingjian. Research on Side Channel Attack Based on CNN-MGU[J]. Netinfo Security, 2022, 22(8): 55-63.

图/表 18

图1

图2

图3

图4

表1

表2

图5

表3

图6

图7

图8

图9

图10

图11

图12

图13

图14

表4

参考文献 18

[1]	KOCHER P, JAFFE J, JUN B. Differential Power Analysis[C]//Springer. Annual International Cryptology Conference. Heidelberg: Springer, 1999: 388-397.
[2]	KOCHER P C. Timing Attacks on Implementations of Diffie-Hellman, RSA, DES, and other Systems[C]//Springer. 16th Annual International Cryptology Conference. Heidelberg: Springer, 1996: 104-113.
[3]	MARTINASEK Z, ZEMAN V. Innovative Method of the Power Analysis[J]. Radio Engineering, 2013, 22(2): 586-594.
[4]	MAGHREBI H, PORTIGLIATTI T, PROUFF E. Breaking Cryptographic Implementations Using Deep Learning Techniques in Security, Privacy, and Applied Cryptography Engineering [C]//Springer. Security, Privacy, and Applied Cryptography Engineering (SPACE 2016). Heidelberg: Springer, 2016: 3-26.
[5]	CAGLI E, DUMAS C, PROUFF E. Convolutional Neural Networks with Data Augmentation against Jitter-Based Countermeasures[C]//Springer. 19th International Conference (CHES 2017), Heidelberg: Springer, 2017: 45-68.
[6]	SAMIOTIS I P. Side-Channel Attacks Using Convolutional Neural Networks: A Study on the Performance of Convolutional Neural Networks on Side-Channel Data[J]. Frontiers in Artificial Intelligence, 2018, 3(1): 73-88. doi: 10.3389/frai.2020.00073 URL
[7]	WANG Kai, YAN Yingjian, GUO Pengfei, et al. Research on Power Analysis Attack Based on Improved Residual Network and Data Augmentation Technology[J]. Journal of Cryptologic Research, 2020, 7(4): 551-564. doi: 10.13868/j.cnki.jcr.000389
	王恺, 严迎建, 郭朋飞, 等. 基于改进残差网络和数据增强技术的能量分析攻击研究[J]. 密码学报, 2020, 7(4): 551-564. doi: 10.13868/j.cnki.jcr.000389
[8]	CHEN Ping, WANG Ping, DONG Gaofeng, et al. SincNet-Based Side Channel Attack[J]. Journal of Cryptography, 2020, 7(5): 583-594.
	陈平, 汪平, 董高峰, 等. 基于SincNet的侧信道攻击[J]. 密码学报, 2020, 7(5): 583-594. doi: 10.13868/j.cnki.jcr.000391
[9]	WANG Junnian, ZHU Bin, YU Wenxin, et al. Side Channel Analysis Based on Deep Learning LSTM[J]. Journal of Computer Engineering, 2021, 47(10): 140-146.
	王俊年, 朱斌, 于文新, 等. 基于深度学习LSTM的侧信道分析[J]. 计算机工程, 2021, 47(10): 140-146.
[10]	WANG Ping, ZHENG Mengce, NAN Jiehui, et al. A Novel Method for Profiling Side Channel Attacks against Imbalanced Data[J]. Journal of Cryptologic Research, 2021, 8(4): 549-559.
	汪平, 郑梦策, 南杰慧, 等. 一种针对侧信道建模攻击中数据不平衡的新方法[J]. 密码学报, 2021, 8(4): 549-559.
[11]	DUAN Xiaoyi, LI You, LINGHU Yunxing, et al. Research on the Method of Side Channel Attack Based on RF Algorithm[J]. Netinfo Security, 2022, 22(1): 19-26.
	段晓毅, 李邮, 令狐韫行, 等. 基于RF算法的侧信道攻击方法研究[J]. 信息网络安全, 2022, 22(1): 19-26.
[12]	CAO Jiahua, WU Zhen, WANG Yi, et al. Side Channel Attack of S-Box Power Randomization Based on CNN-BPR[J]. Journal of Chengdu University of Information Technology, 2022, 37(1): 16-20.
	曹家华, 吴震, 王燚, 等. 基于CNN-BPR的S-Box功耗随机化侧信道攻击[J]. 成都信息工程大学学报, 2022, 37(1): 16-20.
[13]	ZHANG Guanghua, YAN Fengru, ZHANG Dongwen, et al. Insider Threat Detection Model Based on LSTM-Attention[J]. Netinfo Security, 2022, 22(2): 1-10.
	张光华, 闫风如, 张冬雯, 等. 基于LSTM-Attention的内部威胁检测模型[J]. 信息网络安全, 2022, 22(2): 1-10.
[14]	XU Guotian, SHENG Zhenwei. DGA Malicious Domain Name Detection Method Based on Fusion of CNN and LSTM[J]. Netinfo Security, 2021, 21(10): 41-47.
	徐国天, 盛振威. 基于融合CNN与LSTM的DGA恶意域名检测方法[J]. 信息网络安全, 2021, 21(10): 41-47.
[15]	ZHOU Guobing, WU Jianxin, ZHANG Chenlin, et al. Minimal Gated Unit for Recurrent Neural Networks[J]. International Journal of Automation & Computing, 2016(13): 226-234.
[16]	STANDAERT F X, MALKIN T G, YUNG M. A Unified Framework for the Analysis of Side-Channel Key Recovery Attacks[C]//Springer. 28th Annual International Conference on the Theory and Applications of Cryptographic Techniques (EUROCRYPT 2009), Heidelberg: Springer, 2009: 443-461.
[17]	BENADJILA R, PROUFF E, STRULLU R, et al. Deep Learning for Side-Channel Analysis and Introduction to ASCAD Database[J]. Journal of Cryptographic Engineering, 2020(10): 163-188.
[18]	WANG Huanyu, BRISFORS M, FORSMARK S, et al. How Diversity Affects Deep-Learning Side-Channel Attacks[C]//IEEE. 2019 IEEE Nordic Circuits and Systems Conference (NORCAS): NORCHIP and International Symposium of System-on-Chip (SoC). New York: IEEE, 2019:1-16.

数据集	训练样本数量/个	测试样本数量/个	最大抖动
ASCAD	50000	10000	0
ASCAD_desync50	50000	10000	50
ASCAD_desync100	50000	10000	100

网络类型	参数设置	激活函数
输入层	尺寸为700•1	—
卷积层	卷积核为70•1,个数为4	ReLU
BN层	—	—
池化层	均值池化,步长为5,尺寸为5	—
全连接层	—	—
MGU层	神经元数为256	ReLU
Dropout层	Drop_out值为0.5	—
输出层	神经元数为256	Softmax

模型名称	网络层数	参数量/个
LSTM	3	1241856
CNN	13	26806656
CNN-MGU	7	442924

模型名称	训练时间	训练准确率
LSTM	1h:11m:35s	90.0%
CNN	5h:36m:21s	90.0%
CNN-MGU	34m:16s	90.0%