基于自动编码器隐空间分类的建模侧信道分析

Profiled Side-Channel Analysis Based on Autoencoders’Latent Space Classification

侧信道分析是现实世界密码系统的主要威胁之一,建模侧信道分析是一类重要的侧信道分析方法,深度学习技术的引入可拓宽建模侧信道分析的应用场景、提升分析效率.自动编码器(auto-encoder,AE)与变分自动编码器(variational AE,VAE)是被广泛研究的深度学习模型,本文将它们引入建模侧信道分析,提出了基于AE与基于VAE隐空间分类的建模侧信道分析方法,并从生成模型的角度对这两种方法的可行性进行了分析.AE和VAE中间产生的隐空间特征可视为侧信道曲线的低维生成特征,提出的两种方法通过训练相应的AE和VAE来提取出能表征原始侧信道曲线的隐空间特征,并通过VAE探讨了隐空间分布为高斯分布时对建模分析效率的影响.随后提出了三种隐空间特征分类策略:基于欧氏距离的分类策略、基于KL散度的分类策略以及基于支持向量机的分类策略,这些策略可对提取出的隐空间特征进行分类,从而完成建模侧信道分析.在DPAv4.1与增加了高斯噪音的ASCAD数据集上的实验结果表明,基于AE和VAE的建模侧信道分析方法使用三种分类策略的攻击效果均大幅度优于池化模板.从猜测熵的角度看,基于VAE的方法仅需10条DPAv4.1的曲线与1660条加了噪音的ASCAD曲线即可使得两者猜测熵为0,而模板攻击则分别分别需要84条和3899条曲线,效率提升分别达到了88.1%与54.7%.这说明在建模侧信道分析的场景下,VAE有着很好的应用潜能.

Side-channel analysis(SCA) poses great threats to real-world cryptographic systems.Profiled SCA is a major branch of SCA.Applications of deep learning models on the profiled SCA vastly extend its attack surface.As one of the prevailing deep learning models,auto-encoders(AE)and variational auto-encoders(VAE)are introduced into profiled SCA in this paper and the feasibility is illustrated using the generative model in the theory of deeplearning.The latent space features of AE and VAE can be regarded as the generative features of a side-channel trace.This enables AE and VAE as feature extraction methods for side-channel traces,and latent space features are extracted as abstract representations of input traces.By introducing VAE,the impacts on profiled SCA’s efficiency is studied when the latent is in Gaussian distribution.Three latent space classification strategies based on Euclidean distance,KL divergence and support vector machine are then introduced respectively to conduct profiled SCA.Results on the DPAv4.1 dataset and ASCAD dataset with manually-added Gaussian noise shows that,the proposed three classification strategies based on both AE and VAE model have better attack performance than pooled template attack(TA).On the guessing entropy indicator,the VAE uses only 10 traces for DPAv4.1 and 1660 traces for ASCAD with noise to make the guessing entropy of the correct key byte converge to zero.Compared with 84 and 3899 traces presented by pooled TA,the analysis efficiency escalation brought by the proposed strategy reaches 88.1%and 54.7%respectively,which means that VAE has great potential in profiled SCA.