Research on Minimum Receptive Field of Convolutional Neural Networks for Energy Analysis Attack

Research on Minimum Receptive Field of Convolutional Neural Networks for Energy Analysis Attack

연구 분야: Cryptography

학회: International Conference on Digital Forensics and Cyber Crime

Energy analysis attack secret information of cryptographic devices by analyzing the energy leaked during the operation of cryptographic devices. In recent years, CNN-based energy analysis attack is a hot research topic in recent years due to the excellent performance of deep learning convolutional neural networks in processing complex data. In previous CNN-based energy analysis attacks, attackers usually use techniques such as lattice search, random search and Bayesian optimization to carry out the parameter optimization process of convolutional neural networks, and this process consumes a relatively large amount of time. In order to solve this problem, this paper studies the energy cycle characteristics of cryptographic chip, the structural characteristics of convolutional neural network, the optimal convolutional neural network feeling field based on the feeling field, and proposes the minimum feeling field model of convolutional neural network for energy analysis attack, which can derive the minimum feeling field suitable for CNN according to the highest operating frequency of the cryptographic chip, the actual operating frequency, and the collection frequency, from the point of view of the feeling field, and only when the CNN has the minimum feeling field, can it be optimized. The model can be trained successfully only when the CNN's receptive field is larger than this value, thus reducing the search range of the CNN model parameters, and verified on the internationally public ASCAD dataset, and the experiments are consistent with the expected results of the proposed scheme in this paper.