基于降噪自编码器的侧信道攻击预处理方法

A Preprocessing Method of Side Channel Attack Based on Denoising Autoencoder

侧信道分析在硬件安全评估中起着至关重要的作用,而降噪预处理可以去除数据曲线包含的部分噪声,提高攻击成功率.然而,当数据中噪声繁杂且期望预处理前后数据规模不减少时,常规的降噪方式效果较差甚至无效.本文基于卷积神经网络设计了一种优化的降噪自编码器.首先,对第一轮加密的字节代换操作具有相同输出的数据曲线做均值滤波处理,并根据字节代换的输出构造对应的自编码器模型标签,最大化地提取出纯净数据.其次,在计算标签与预测值的损失函数中添加L2正则化惩罚项,防止过拟合以及加速训练.本文对公开的DPAContestV2、DPAContestV4.1和ASCAD数据集进行降噪预处理及侧信道攻击.实验结果表明,处理后的数据相比原始数据信噪比分别提高3.53、3.14、3.86倍,皮尔逊相关系数分别提高1.94、1.37、1.04倍.在攻击阶段,不进行降噪预处理时V2、V4.1、ASCAD数据集分别需要1175、4、191条测试轨迹破译密钥.而使用本文方法降噪后成功攻击所需轨迹数量分别降低为440、1、41条.因此,本文的降噪自编码器网络可以大幅度降低信号中包含的噪声,并显著提高了攻击性能.

Side channel attack plays a vital role in hardware security evaluation,and noise reduction preprocessing can remove part of the noise contained in traces and improve the probability of successful attacks.However,since electronic noise is diverse and the number of available traces does not decrease significantly due to denoising,the known noise reduction methods do not work well.This paper designs an optimized denoising autoencoder based on a convolutional neural network.First,this paper applies mean filtering on the original traces which have the same output after a SubBytes operation in the first round of encryption and constructs,and constructs the label of the corresponding autoencoder model to extract pure data.Then the L2 regularization penalty term is applied to the loss function between the label and the prediction value to prevent overfitting and accelerate the training process.In this paper,the public datasets DPA Contest V2,DPA Contest V4.1,and ASCAD data sets are denoised and side-channel attacks are carried out.The experimental results show that the signal-tonoise ratios of processed data are increased to 3.53,3.14,and 3.86 times respectively compared to the original data,and the Pearson correlation coefficients are increased by 1.94,1.37,and 1.04 times respectively.Moreover,if no denoising preprocessing is performed,1175,4,and 191 traces are required to recover the secret key separately for the V2,V4.1,and ASCAD datasets.However,the number of traces required for a successful attack is reduced to 440,1,and 41 respectively.Thus,the denoising autoencoder network proposed in this paper can greatly reduce the noise contained in the traces and significantly improve the performance of side-channel attacks.