Deep-Learning Based Nonprofiling Side-Channel Attack on Mask Leakage-Free Environments Using Broadcast Operation

With the recent development of artificial intelligence (AI), efforts to apply related technologies to various fields are rapidly increasing. In the field of cryptanalysis, research utilizing deep learning is continuously being published in order to keep up with this trend. Side-channel analysis is a a type of cryptanalysis that uses physical information and can be classified into profiling and nonprofiling analyses. Nonprofiling attacks using deep learning take advantage of the fact that training is performed relatively well when the right key is compared to the wrong key. Masking countermeasures are applied to design a secure cipher against side-channel analysis. The traditional second-order attack for analyzing masked ciphers is used by preprocessing the side channel information to remove the mask value. However, deep learning has the advantage of being able to omit this process. Related works proposed so far attempted to analyze the masked cipher, but focused only on 1-byte analysis using the masking information itself. In reality, grasping the time-points, in which only the masking information is revealed, is difficult and far from the secret key analysis area. In this study, we attempt to analyze the case of combining masked 2-byte information, not only using the masking information. We also propose a neural network design scheme to perform more effective attacks. The proposed method highlights the relative difference between the right and wrong keys. Previous research on analysis evaluation criteria has been lacking. Therefore, we propose herein new evaluation metrics that can be easily used and demonstrate their validity using simulation and actually collected data. As a result of the experiment, the proposed methods based on the loss metric improved by approximately 228.59% in the simulation dataset and 739.46% in the real dataset compared to the binary labeling. And it reduced the minimum number of analytical traces by approximately 78.95% and 72.5%, respectively