关键词: 侧信道攻击, 自动编码器, 信息安全, 去噪与降维, 数据预处理

Abstract: To address the reduction of signal-to-noise ratio and feature separability of leakage traces in side channel attack (SCA) under strong noise and hiding countermeasures, which significantly increases the number of traces required for guessing entropy convergence, a side channel attack preprocessing method for high-order neighbor-constrained autoencoder, termed N-CAE (Neighbor-Constrained Autoencoder), was proposed. First, a high-order adjacency matrix was constructed in the original trace space and combined with label consistency to form a structure-guided matrix. Then, a neighbor loss function was introduced into a convolutional autoencoder framework and jointly optimized with the reconstruction loss. Next, structural constraints were employed to regulate the distribution of low-dimensional features produced by the encoder, enabling intra-class aggregation and inter-class separation. Finally, 512-dimensional latent representations were extracted as inputs to the attack model, achieving collaborative optimization of denoising and dimensionality reduction. Experiments were conducted on the ASCAD fixed-key dataset under five typical interference scenarios, including Gaussian noise, desynchronization, random delay interruption, clock jitter, and shuffle, show that N-CAE outperforms CAE (Convolutional Autoencoder), D-CAE (Dilated Convolutional Autoencoder), and DAE (Denoising Autoencoder) in signal-to-noise ratio (SNR), Pearson correlation coefficient (PCC), and attack performance (guessing entropy). In the shuffle scenario, the peak SNR reaches 0.39, representing an improvement of 1850% compared with the original traces and 358.9% over the best competing method CAE. In the Gaussian noise scenario, the number of traces required to achieve guessing entropy convergence is reduced by 60.62%. The proposed method improves trace signal quality and feature separability under multiple hiding countermeasures, thereby significantly enhancing attack performance.

Key words: Side Channel Attack (SCA), Autoencoder, Information security, Denoising and dimensionality reduction, Data preprocessing