基于深度残差神经网络的AES密码芯片电磁攻击研究

基于"分而治之"方法提出了一种在完全未知明文、密文及泄露中间值情况下的电磁泄漏攻击方法.设计深度残差神经网络模型，对基于现场可编程逻辑门阵列（field programmable gate array，FPGA）的密码芯片高级加密标准（advanced encryption standard，AES）加密算法进行了电磁分析攻击.该模型包括数据扩展层和深度残差层两部分.数据扩展层将一维电磁信号数据扩展到二维，有效降低了模型的训练难度；深度残差层是基于残差块的深度神经网络，有效解决了深层网络的收敛难、调优难等问题.在明文和密文完全未知的情况下，仅仅通过采集到的电磁泄漏信号，利用该模型对密钥的最后两位进行了恢复实验，实验结果表明准确率达到了91.8%.在同等条件下，该模型的准确度比支持向量机（support vector machine，SVM）模型提升了近8%.

EM attack on AES cryptography chip based on deep residual neural network

An electromagnetic attack method is proposed in the case of completely unknown plaintext, cipher text and leakage intermediate value. In this paper, a deep residual neural network model is designed to achieve electromagnetic attack on advanced encryption standard(AES) encryption algorithm based on field programmable gate array(FPGA). The model consists of two parts, the data expansion layer and the depth residual layer. The data expansion layer extends the electromagnetic signal data from one dimension to two dimensions, effectively reducing the training difficulty of the model. The deep residual layer is a deep neural network based on residual blocks, effectively solving the problems of the difficulties of deep network convergence and difficulty in tuning. The experimental results show that the final two bits of the key can be obtained only through the collected electromagnetic leakage signals, and the accuracy rate is 91.8%. Under the same conditions, the accuracy of the model is nearly 8% higher than that of the support vector machine(SVM) model.