SCARefusion: Side channel analysis data restoration with diffusion model

Abstract

Side channel Analysis (SCA) based on deep learning is highly sensitive to data quality. When using the ID leakage model as the labeling criterion in supervised classification problems, slight data imbalance issues arise, which can reduce analysis efficiency. Diffusion Models are an emerging class of generative models that offer more intuitive, stable, robust, and interpretable advantages compared to Generative Adversarial Network (GAN). To address data imbalance, we introduce a latent diffusion model based on U-Net that retains high-resolution information for the decoding process, thereby recovering detailed power consumption trace (hereinafter referred to as traces) information. Our model, named SCA Restoration with Diffusion Model (SCARefusion), comprises Conditional Nonlinear Activation Free Blocks (CNAFBlocks), downsampling, and upsampling modules. The network integrates the SimpleGate nonlinear activation function, enhancing model performance and computational efficiency, and improving adaptability to input data. This approach effectively generates balanced class data for labels, mitigating dataset imbalance and avoiding the instability of GAN training. In this paper, Correlation Power Analysis (CPA) is used to calculate and compare the correlation coefficients between hypothetical and measured traces to detect whether the generated data exhibit the same leakage points as the original data. Additionally, the effectiveness of the generated traces is validated using Convolutional Neural Network (CNN) attacks. Experimental results demonstrate that SCARefusion exhibits outstanding performance on the fixed-key ASCAD synchronous and asynchronous 50 dataset, capable of generating traces consistent with the leakage points of the original traces and successfully extracting the correct key. Furthermore, we discuss reducing the learning rate, optimizing network architecture, and other parameters to address loss oscillation issues during model training. These studies contribute to improving the performance and robustness of deep learning in SCA, effectively addressing data challenges.