文献类型：期刊文献

中文题名：基于多尺度特征融合的功耗建模分析方法

英文题名：Power profiling analysis method based on multi-scale feature fusion

作者：李想[1];杨宁[1];刘伟锋[2];陈艾东[3,4];张彦龙[5];王硕[5];周婧[5]

第一作者：李想

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]中国科学院半导体研究所,北京100085;[3]北京联合大学机器人学院,北京100101;[4]北京联合大学多智能体系统研究中心,北京100101;[5]北京微电子技术研究所,北京100076

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2024

卷号：61

期号：3

起止页码：182-194

中文期刊名：四川大学学报（自然科学版）

外文期刊名：Journal of Sichuan University(Natural Science Edition)

收录：CSTPCD;;北大核心:【北大核心2023】；CSCD:【CSCD_E2023_2024】；

基金：国家重点研发计划项目(2022YFB2804402)。

语种：中文

中文关键词：侧信道分析;功耗分析;多尺度特征融合;离散小波变换;注意力机制

外文关键词：Side-channel analysis;Power analysis;Multi-scale feature fusion;Discrete wavelet transform;Attention mechanism

摘要：在数字化浪潮推动下,5G和6G技术的快速发展正引领移动通信系统步入新阶段.先进的硬件设备和加密芯片为不断增长的数据处理需求和日益关注的安全保障提供了强有力的支持.在这一背景下,搭载现代密码技术的各类硬件设备逐渐演变为不可或缺的生活基石.这些设备已经具备抵御传统密码分析的能力.近年来,学术界的研究重点之一是对设备在实际运行过程中产生的物理泄漏进行分析.这一领域被称为侧信道分析(Side-Channel Analysis,SCA).深度学习驱动的侧信道分析已被广泛认可为一种有效的方法,针对当前神经网络模型的功耗曲线数量需求大、鲁棒性差和收敛速度慢等问题,本文提出一种基于CNNbest的多尺度特征融合侧信道分析方法.首先,重构特征提取网络结构,以解决深层特征向量容易过度解释噪声细节的问题和模型过拟合问题.而后,使用滤波器阵列执行离散小波变换(Dis‐crete Wavelet Transform,DWT)分析方法构造多解析度时频,提升数据质量.最后,引入轻量级的结合通道空间的卷积块注意力模块(Convolutional Block Attention Module,CBAM),以提高功耗曲线关键特征的学习效率.实验结果表明,本文方法对侧信道分析所需的功耗曲线较原模型减少了88.27%,显著提高了分析性能,能够满足侧信道建模和分析的要求.
In the wave of digitization,the rapid development of 5G and 6G technologies is leading the mobile communication systems into a new era.Advanced hardware devices and encryption chips offer robust support for the escalating demand in data processing and the growing emphasis on security.In this context,various hardware devices equipped with modern cryptographic technology are gradually evolving into indispensable cornerstones of our daily lives.These devices have the capability to resist traditional cryptographic analysis.Inrecent years,one of the focuses of academic research is the analysis of physical leakage occurring during the actual operation of devices,a field known as Side-Channel Analysis(SCA).Deep learning-driven side-channel analysis has been widely recognized as an effective method.Aiming at the current neural network model's problems such as high demand for the number of traces,poor robustness,and slow convergence speed,this paper proposes a multiscale feature fusion side-channel analysis method based on CNNbest.Firstly,the structure of the feature extraction network is revised to mitigate the issue of deep feature vectors being susceptible to excessive interpretation of noise details and model overfitting.Subsequently,a filtering array is used to perform Discrete Wavelet Transform(DWT)analysis,constructing multi-resolution time-frequency representations to enhance data quality.Finally,a lightweight Convolutional Block Attention Module(CBAM)incorporating channel spatial attention is introduced to improve the learning efficiency of key features in power consumption curves.Experimental results demonstrate that the proposed method reduces the power consumption curves required for side-channel analysis by 88.27%compared to the original model,significantly improving analysis performance and meeting the requirements of side-channel modeling and analysis.