Abstract

In recent years, deep learning has been actively applied to the field of side-channel analysis, and has focused on profiling scenarios in particular. The advent of the Differential Deep Learning Analysis (DDLA) technique allows the advantages of deep learning to be utilized in non-profiling scenarios, where for a fixed key, only a limited number of power traces can be analyzed. However, in most DDLA-related studies, training metric graphs for estimating performance are only provided, without specific numerical value. In this case, there are several difficulties, such as performance comparison and implementation of automatic attack scripts. In this paper, we propose a novel performance metric, the Normalized Maximum Margin (NMM), which takes into account the statistical characteristics of training metric values, such as accuracy and loss value. In addition, we present three experiments to verify that the NMM can be effectively used for attack success decision, performance evaluation, and so on. These experimental results show that the NMM can objectively represent the performance of DDLA, and whether the attack was successful. Furthermore, we confirm that the NMM can also be used as a key distinguisher, which allows the key to be extracted in the real world, where the correct key is unknown.