基于降噪自编码器的侧信道攻击预处理方法OA北大核心CSTPCD

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-to-noise 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.