LightSCA: Lightweight Side-Channel Attack via Discrete Cosine Transform and Residual Networks
Nengfu Cai, Du Wang, Md Zakirul Alam Bhuiyan, Lihua Han, Gang Li · 2022
Side-channel attacks (SCAs) crack the keys through physical signals that are unintentionally leaked during the operation of cryptographic devices. This poses a severe threat to users' privacy and network security. Deep learning (DL)-based SCAs have become one of the most important means of evaluating the security of cryptographic devices. However, the existing SCAs still have the following limitations. First, the existing SCAs based on convolutional neural networks (CNNs) have large networks sizes and take a long time to train the model. Second, sensitivity to desynchronized signals is also a significant problem. For desynchronized signals, the side channel traces required to crack the key increase rapidly compared to the synchronized signals. In this background, we analyze these problems in detail and aim to design a SCA algorithm that is lightweight and robust to desynchronized signals. In this paper, we propose a Lightweight Side-Channel Attack model, named LightSCA, based on discrete cosine transform (DCT) and residual networks. We use the DCT to extract the pivotal features of the side channel traces. Specifically, the feature length is compressed to 1/35 of the original. In order to improve the efficiency of model training and key cracking, we design a multi-classifier based on residual networks. The number of LightSCA network parameters is less than 1/100 of the CNNbest model. We validate the performance of LightSCA on the public ASCAD dataset. Compared with four representative DLSCA models, the proposed model achieves better effectiveness. For the synchronization signals, LightSCA cracks a key byte using only 87 traces.