Breaking Hardware Implementation of SM4 Algorithm using Principal Component Analysis and Machine Learning Classifiers

In this paper, we executed a power analysis of hardware implementation of SM4 algorithm using machine learning classifiers (support vector machine (SVMs), Decision Trees (DT), k-Nearest Neighbor, (KNN), Ensemble learning (EB), etc.). We first download hardware design to the FPGA in power acquisition platform based on SASEBO-G board to obtain the power traces of SM4 during its encryption process. Then we introduced a machine learning method to the Hamming-weight attack model of SM4 algorithm. In the data preprocessing phase, principal component analysis (PCA) method was applied, and we explored how many points of interested (PoIs) should we take to achieve the best test accuracy. We have tried different kinds of classifiers and their performance against Gaussian noise. The results show that for different classifiers, the number of PoIs selected when achieving maximum precision is different, PCA can reduce the dimensions of power trace effectively but probably decrease the test accuracy. Furthermore, most of these classifiers have excellent resolution (up to 100%) for Gaussian noise superimposed on the power traces, which means that machine learning classifiers such like SVM and Decision Trees can compete with template attacks based on Gaussian distribution.