A nonprofiled side-channel analysis based on variational lower bound related to mutual information

In this paper, we attempt to improve the practical performance of the nonprofiled side-channel analysis (NonSCA) with the help of neural networks. We first derive a variational lower bound related to mutual information (VLBRMI) optimized for the context of NonSCA, which possesses a set of adjustable parameters and whose maximum value linearly depends on the mutual information. Then, we propose a new NonSCA method called neural mutual information analysis (NMIA) that exploits the maximum VLBRMI as the distinguisher. We present an estimator of the maximum VLBRMI, which uses neural networks to instantiate the VLBRMI and trains the neural networks to approximate the maximum VLBRMI so that we can implement the NMIA efficiently. Finally, we evaluate the NMIA on several datasets. The experimental results show that NMIA outperforms the correlation power analysis, the mutual information analysis (MIA) based on histograms, the MIA based on kernel density estimation, and the state-of-the-art NonSCA method based on neural networks.