背景:基于卷积神经网络(CNNSCA)的侧通道密码分析方法可以有效地进行密码攻击。实现密码分析的CNNSCA网络模型主要包括基于VGG变体的CNNSCA(VGG-CNNSCA)和基于Alexnet变体的CNNSCA(Alex-CNNSCA)。这些CNNSCA模型的学习能力和密码分析性能并不理想,并且训练后的模型精度较低,训练时间太长,并占用更多的计算资源。为了提高CNNSCA的整体性能,本文将改进CNNSCA模型设计和超参数优化。
方法:本文首先研究了SCA应用场景中的CNN架构组成,并推导了CNN核心算法对一维数据侧信道泄漏的计算过程。其次,综合运用VGG-CNNSCA模型分类、拟合效率和Alex-CNNSCA模型占用计算资源少的优势,设计了一种新的CNNSCA基础模型,为了更好地减少深度网络中误差反向传播的梯度色散问题,SE(挤压和激励)模块新嵌入到这个基本模型中,该模块首次在CNNSCA模型中使用,形成了CNNSCA模型设计的新思路。然后将此基本模型应用于来自边信道泄漏公共数据库(ASCAD)的已知一阶屏蔽数据集。在此应用场景中,根据模型设计规则和实际实验结果,排除非必要的实验参数。在最客观的实验参数区间内优化基本模型的各种超参数,提高其密码分析性能,这导致了超参数优化方案和确定超参数的最终基准。
结果:最后,获得了一种新的CNNSCA模型优化架构,用于攻击未受保护的加密设备-CNNSCAnew。通过对比实验,CNNSCAnew的猜测熵评价结果收敛到61。从模型训练到成功恢复的关键,总时间缩短到30分钟左右,我们获得了比其他CNNSCA模型更好的性能。
BACKGROUND: The side-channel cryptanalysis method based on convolutional neural network (CNNSCA) can effectively carry out cryptographic attacks. The CNNSCA network models that achieve cryptanalysis mainly include CNNSCA based on the VGG variant (VGG-CNNSCA) and CNNSCA based on the Alexnet variant (Alex-CNNSCA). The learning ability and cryptanalysis performance of these CNNSCA models are not optimal, and the trained model has low accuracy, too long training time, and takes up more computing resources. In order to improve the overall performance of CNNSCA, the paper will improve CNNSCA model design and
hyperparameter optimization.
METHODS: The paper first studied the CNN architecture composition in the SCA application scenario, and derives the calculation process of the CNN core algorithm for side-channel leakage of one-dimensional data. Secondly, a new basic model of CNNSCA was designed by comprehensively using the advantages of VGG-CNNSCA model classification and fitting efficiency and Alex-CNNSCA model occupying less computing resources, in order to better reduce the gradient dispersion problem of error back propagation in deep networks, the SE (Squeeze-and-Excitation) module is newly embedded in this basic model, this module is used for the first time in the CNNSCA model, which forms a new idea for the design of the CNNSCA model. Then apply this basic model to a known first-order masked dataset from the side-channel leak public database (ASCAD). In this application scenario, according to the model design rules and actual experimental results, exclude non-essential experimental parameters. Optimize the various hyperparameters of the basic model in the most objective experimental parameter interval to improve its cryptanalysis performance, which results in a hyper-parameter optimization scheme and a final benchmark for the determination of hyper-parameters.
RESULTS: Finally, a new CNNSCA model optimized architecture for attacking unprotected encryption devices is obtained-CNNSCAnew. Through comparative experiments, CNNSCAnew\'s guessing entropy evaluation results converged to 61. From model training to successful recovery of the key, the total time spent was shortened to about 30 min, and we obtained better performance than other CNNSCA models.