Template Attacks Based on Two-dimensional Convolutional Neural Networks

In the TA(Template Attack) of the side channel attack, the template construction method has been developed from the original Gaussian distribution model to the current machine learning model, such as SVM(Support Vector Machines), CNN(Convolutional Neural Networks), and so on. The traces has not been pre-processed which contains noise, jitter and other interference factors, 1D-CNN(One-dimensional Convolutional Neural Networks) often require many energy consumptions curves to train templates with an attack success rate of more than 90%. This paper proposes a new type of 2D-CNN(Two-dimensional Convolutional Neural Network) structure for template attacks based on the existing CNN1D. The premise of this network structure is to perform SFFT(Sparse Fast Fourier Transform) on the data set, and then, train the pre-processed multi-dimensional data through a CNN2D. Result shows that the CNN2D uses a template trained with less traces to achieve the same attack success rate as the original CNN1D.