Profiling side-channel attacks based on CNN model fusion

Abstract

The deep learning algorithm analyzes the power consumption, electromagnetic, and other side-channel information leaked by hardware devices, which has powerful attack capability. However, there is also the problem of high difficulty in hyperparameter tuning. To this end, we propose a side-channel attack scheme using convolutional neural network (CNN) model fusion. First, N sets of hyperparameters were randomly selected in the search space, and the CNN model was trained using the side-channel dataset to obtain N base models. Secondly, the high-dimensional features in the middle layer of the base model were merged to increase the amount of effective information and enhance the generalization ability of the neural network model, and a new fusion model was constructed by training. Finally, the fusion model was used to predict the output probability of DPA Contest v4, ASCAD, AES_HD, and AES_RD public side-channel datasets, analyze the guessing entropy of each key to recover the key, and evaluate the attack efficiency of the scheme. The experimental results show that the CNN model fusion method can effectively improve the side-channel attack capability and reduce the difficulty of hyperparameter tuning.