Power Analysis based on Deep Learning for Multiple Implementations of Cryptographic Algorithms

Power analysis is a side-channel method widely studied and proven effective in attacks. Deep learning-based power analysis can extract higher-level and more abstract features from power traces, enabling attacks on cryptographic devices with protective measures. There are many types of neural networks used in deep learning, and this paper applies various neural networks to power analysis of Advanced Encryption Standard (AES) algorithm implementations on Microcontroller Unit (MCU) and Field Programmable Gate Array (FPGA). The results show that for implementations of cryptographic algorithms with no or weak protection, Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and CNN-LSTM networks all demonstrate good attack effectiveness. When the training set is insufficient, CNN and CNN-LSTM perform better. For unsecured or first-order fixed mask-implemented cryptographic algorithms, simple network structures can also achieve good attack results.