EM Attack On FPGA Cryptography Chip Based On Deep Neural NetWork

Cryptographic chip's electromagnetic analysis attack refers to an attack method that uses the electromagnetic leakage signal of the cryptographic chip to obtain the key-related information in the encryption algorithm.Because the electromagnetic leakage signal has the advantages of abundant information and no contact for collection,the research on electromagnetic analysis attacks is of great significance for the design and construction of high security cryptographic chips.In this paper,an electromagnetic analysis attack is performed on the FPGA chip running the AES encryption algorithm.Without clear text and encryption intermediate value,the attack is directly targeted at the 128-bit secret key of AES.The main tasks are as follows.1.In order to solve the limit of the data set,a data enhancement scheme is proposed to generate a large amount of data.This scheme not only solves the problem of neural network overfitting caused by too small data set,but also greatly improves the signal-to-noise ratio of the superimposed data,which is more conducive to extract the effective features with the neural network.The experiment proves that the test accuracy of the neural network trained using the superimposed data set is close to 40 percentage points higher than that of the original data set.2.A convolutional residuals block containing an approximate identity connection(A-Identity)structure named CIResBlock,is proposed and designed for different residuals of input and output,and the residuals neural network structure is designed based on this.Combining the idea of "divide and conquer",the classification model of residual neural network was constructed for the 128-bit secret keys of AES,which were divided into 32 groups according to each 4-bit group.Experiment shows that the accuracy of these 32 groups of network models are above 80%,and the highest reaches 85.2%.Compared with the Inception V2 network model used in reference[1],the accuracy is nearly 13 percentage points higher.Through comparative experiments,the accuracy of the network model with A-Identity structure is 4 percentage points higher than that without a-identity structure,and the convergence rate is faster.3.In order to extract the sequential correlation features in the electromagnetic leakage signal,a residual Recurrent neural network is designed,which uses the residual convolution layer of the 2nd innovation to extract the local feature sequence of the signal,then inputs the feature sequence into the LSTM network to extract the sequential correlation features,and finally makes classification.In this network structure,in order to make more effective use of the timing correlation features of LSTM data,we have designed an LSTM-FC structure which uses the feature values of the LSTM at each moment.Similarly,32 models were trained respectively with this network structure.Compared to the network model in the 2nd innovation,its accuracy was 3 to 5 percentage points higher on average,and the highest reached 89.2%.