基于CBAPD网络的侧信道攻击

Side-channel Attacks Based on CBAPD Network

侧信道攻击是一类强大的密码分析攻击,自该理论提出以来受到了密码学界的广泛关注.近年来深度学习技术被越来越多地应用于侧信道攻击领域,其中如何提升深度学习模型的性能是研究的热点.本文根据攻击目标数据的特点,提出了一种新的卷积神经网络结构CBAPD,此网络将卷积层中的激活函数去除,然后在卷积层后加入了批标准化层,并且在批标准化层后加入一个激活层来激活敏感信息.为评估模型的性能,在两个公开数据集ASCAD和DPA-contestv4上进行了测试.实验结果表明,本文所提出的CBAPD网络在ASCAD同步数据集上仅需要50条能量迹就可以攻击成功,在最大异步量为50和100个样本点的数据集上分别需要160和1850条能量迹就可以使rank值降到0并保持不变.在DPA-contestv4数据集上,CBAPD模型仅需要3条能量迹即可攻击成功.同时,通过对比2019年Benadjila等人所提出的CNN_(best),2020年陈等人所提出的SincNet网络和Zaid等人所提出的模型,CBAPD模型在最大异步量为50个样本点的ASCAD数据集上成功攻击时所需能量迹可减少34.426%-96.8%.而在DPA-contestv4数据集上,CBAPD模型与Zaid等人所提出的模型攻击效果相同,且优于其他两个模型.因此,本文所提出的CBAPD模型在不同的数据集上均有良好的表现.

Side-channel attacks are powerful cryptanalytic attacks,which have attracted extensive attention in the society of cryptography since their proposal.In recent years,deep learning technique has been increasingly applied to side-channel attacks,and how to improve its performance is a hot spot of research.Based on characteristics of the target data,a new convolutional neural network,named CBAPD,is presented.The new network removes the activation function in a basic convolutional layer,adds a batch normalization layer after the convolutional layer,and then adds an activation layer after the batch normalization layer to activate sensitive information.To evaluate performance of the CBAPD,two public datasets,ASCAD and DPA-contest v4,are tested.Experiment results show that the proposed network needs only 50 traces on the ASCAD synchronous dataset for successful attacks,and requires 160 and 1850 traces on the datasets with the maximum asynchronous sizes of 50 and 100 sample points respectively,to make the rank value drop to 0 and then remain unchanged.On the DPAcontest v4 dataset,the CBAPD model needs only 3 traces to lunch an attack successfully.Meanwhile,compared with the CNNbest model proposed by Benadjila et al.in 2019,the SincNet proposed by Chen et al.and the model proposed by Zaid et al.in 2020,the traces required by CBAPD model for successful attacks on the ASCAD dataset with the maximum asynchronous sizes of 50 sample points can be reduced by 34.426%∼96.8%.On the DPA-contest v4 dataset,the CBAPD model has the same attack effect as the model proposed by Zaid et al.and outperforms the other two models.Therefore,the proposed CBAPD model has good performance on different datasets.