| 引用本文： |
-
赵竟霖,张倩,邱心宽,周永彬,乔泽华.面向基于深度学习侧信道分析的通用数据增强方法[J].信息安全学报,已采用 [点击复制]
- zhao jing lin,zhang qian,qiu xin kuan,zhou yong bin,qiao ze hua.A General Data Augmentation Method for Deep Learning Based Side-Channel Analysis[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 数据增强方法可以在不增加实际采集的训练样本的情况下，构建更加准确的深度学习模型，从而提高基于深度学习的侧信道分析的技术效果。然而，当前已有数据增强方法的参数选取通常高度依赖专家知识，且只适合对特定算法实现数据集进行数据增强。因此，构造一个适用于多种算法实现数据集的通用数据增强方法、自适应地选取数据增强参数，具有重要现实意义。本文提出了一种通用数据增强方法，其核心思想是采用模拟退火机制依据反馈信号自适应地选择数据增强策略参数。该方法主要由三个核心组件构成：采用模拟退火机制的主控制器、采用组合式数据增强机制的数据增强单元以及提供侧信道攻击代价反馈信号的攻击评估单元。以无防护软件和硬件AES实现、随机延迟和掩码防护AES软件实现为分析目标，在相应的本领域公开数据集上，开展了基于深度学习的侧信道分析实验对比研究，验证了本文所提出通用数据增强方法的有效性。具体地，在AES_HD、DPA v4、ASCADf（N=0/N=50/N=100）三种数据集五个场景下，与未采用数据增强方法的DL-SCA效果相比，攻击成功时所使用的最少能量迹条数分别降低21%、25%、24%、48%以及5%。在AES_HD、DPA v4、ASCADf（N=0/N=50）三种数据集四个场景中，与采用其他数据增强方法的DL-SCA效果相比，攻击成功时所使用的最少能量迹条数分别降低22%、40%、10%以及30%。 |
| 关键词: 深度学习 侧信道分析 数据增强 自适应参数选取 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2025.04.23 |
| 投稿时间：2024-02-09修订日期：2024-03-20 |
| 基金项目:本文工作得到国家重点研发计划(No. 2022YFB3103800)、国家自然科学基金(No.U2336205、No.62202231、No.62202230、No.62302224、No.62302226)、中国博士后科学基金(No.2023M741709)、江苏省卓越博士后计划(No.2023ZB031)、云南省重大科技专项计划（NO.202302AD080002)资助。 |
|
| A General Data Augmentation Method for Deep Learning Based Side-Channel Analysis |
|
zhao jing lin, zhang qian, qiu xin kuan, zhou yong bin, qiao ze hua
|
| (Institute of Information Engineering, Chinese Academy of Sciences) |
| Abstract: |
| Data augmentation methods can construct more accurate deep learning models without increasing the number of acquired training samples, thereby improving the technical effectiveness of deep learning based side-channel analysis. Nonetheless, the selection of parameters for current data augmentation techniques is predominantly dependent on specialized expertise and is tailored specifically to datasets associated with certain cryptographic algorithm implementations. Consequently, developing a universal data augmentation strategy that is versatile across various cryptographic algorithm implementations and capable of adaptively choosing augmentation parameters holds significant practical value. In this study, we introduce a comprehensive data augmentation approach designed to address this challenge. This method is structured around three core components: a controller employing the simulated annealing algorithm, a data augmentation module that leverages a sophisticated combination of augmentation techniques, and an attack evaluation module that generates feedback in the form of cost metrics. We validated the efficacy of this innovative approach through comparative experiments on widely recognized datasets in the domain of side-channel analysis, focusing on unprotected software and hardware AES implementations, as well as AES software implementations fortified with random delay and masking techniques. Our analysis, utilizing deep learning for side-channel attack investigation, confirmed the superior performance of our proposed universal data augmentation method across various scenarios. Specifically, in the case of the AES_HD, DPA v4, and ASCADf datasets, across five different settings with N=0, N=50, and N=100, we observed reductions in Measurements To Disclosure (MTD) by 21%, 25%, 24%, 48%, and 5%, respectively, when compared to DL-SCA approaches without data augmentation. Moreover, in four distinct scenarios across the AES_HD, DPA v4, and ASCADf datasets with N=0 and N=50, the MTD was reduced by 22%, 40%, 10%, and 30%, respectively in comparison to DL-SCA with SOTA data augmentation methods. |
| Key words: deep learning side-channel analysis data augmentation adaptive parameter selection |
