Convolutional Neural Networks for Profiled Side-Channel Analysis

Recent studies have shown that deep learning algorithms are very effective for evaluating the security of embedded systems. The deep learning technique represented by Convolutional Neural Networks (CNNs) has proven to be a promising paradigm in the profiled side-channel analysis attacks. In this paper, we first proposed a novel CNNs architecture called DeepSCA. Considering that this work may be reproduced by other researchers, we conduct all experiments on the public ASCAD dataset, which provides electromagnetic traces of a masked 128-bit AES implementation. Our work confirms that DeepSCA significantly reduces the number of side-channel traces required to perform successful attacks on highly desynchronized datasets, which even outperforms the published optimized CNNs model. Additionally, we find that DeepSCA pre-trained from the synchronous traces works well in presence of desynchronization or jittering after a slight fine-tuning.