An Optimized Non-profiled Deep Learning-Based Power Analysis with Self-supervised Autoencoders

With the rapid development of deep learning, it has been adopted as a primary method of analysis in non-profiled side-channel attacks. However, due to the noises in the collected power traces and the significant amount of data required to train a deep learning neural network, the non-profiled deep learning analysis method faces challenges in practical application. In this paper, a novel non-profiled differential deep learning analysis architecture that incorporates a self-supervised autoencoder is proposed. The autoencoder is designed to reduce the noise and strengthen the features of power traces before they are used as training data for the neural network. The experiment results indicate that not only the architecture outperforms the traditional differential deep learning network with more distinction, but it also distinguishes the correct key with a lower computational cost. The architecture is also examined with small datasets and is proved to be able to maintain the capability of recovering the correct key when the traditional architecture has failed.