基于半监督学习的AES算法功耗分析  

Power Analysis Attack on AES Based on Semi-Supervised Learning

作　　者：王相宾 王永娟[1,2] 赵远 高光普[1,2] 袁庆军 WANG Xiang-Bin;WANG Yong-Juan;ZHAO Yuan;GAO Guang-Pu;YUAN Qing-Jun(PLA Strategic Support Force Information Engineering University,Zhengzhou 450001,China;Henan Key Laboratory of Network Cryptography,Zhengzhou 450001,China)

机构地区：[1]战略支援部队信息工程大学,郑州450001 [2]河南省网络密码重点实验室,郑州450001

出　　处：《密码学报》2021年第4期660-668,共9页Journal of Cryptologic Research

基　　金：国家自然科学基金(61402522);河南省网络密码技术重点实验室开放基金(LNCT2019-S02)。

摘　　要：基于机器学习的功耗分析是目前功耗分析的主要研究方向之一,属于建模类的攻击.针对无掩码防护的AES算法实现,本文将半监督机器学习算法Tri-Training应用于功耗分析,有效减少了用机器学习算法进行建模时所需要的有标记能量迹数量.相较于基于有监督机器学习的建模类功耗分析,使用Tri-Training算法可以有效减小对有标记能量迹的需求,更具有现实意义.然而,Tri-Training算法在初始分类器较弱时,容易出现错误标记现象,影响分类的准确率和建模的效率.对此本文在使用Tri-Training算法进行建模时引入了阈值判断操作,提高了分类的准确率,并对比了不同阈值对分类准确率的影响.本文对在ATM89S52单片机上实现的AES-128算法进行建模类功耗分析,实验结果表明,在使用80条有标记能量迹时,相较于使用有监督学习算法的准确率为63.49%,本方法的准确率为74.56%,准确率提升了约11%.Power analysis attack based on machine learning is currently one of the main research directions of power consumption analysis,which belongs to profiling attacks.In this paper,a power analysis model establishment method based on the semi-supervised machine learning algorithm Tri-Training is proposed for the AES algorithm without masked protection,which effectively reduces the number of labeled power traces required for profiling attacks based on machine learning.Compared with the profiling power analysis based on supervised machine learning,profiling power analysis based on the Tri-Training algorithm can greatly reduce the demand for labeled power traces.However,the Tri-Training algorithm is prone to mislabeling when the initial classifier is weak,which affects the accuracy of classification and the efficiency of model establishment.In this regard,this paper introduces a threshold judgment operation when using the Tri-Training algorithm for modeling to improve the accuracy of classification,and compares the impact of different thresholds on the accuracy of classification.This paper conducts experimental analysis on the AES-128 algorithm implemented on ATM89S52 single-chip microcomputer.The experimental results show that,when 80 labeled power traces are used,the accuracy rate is 74.56%,while it is 63.49%using the supervised learning algorithm.The accuracy rate has increased by about 11%.

关 键 词：功耗分析 半监督学习 Tri-Training算法 AES-128算法 能量迹 

分 类 号：TN918.1[电子电信—通信与信息系统]