Paper 2025/1697

Extract Discriminative Features: Profiled Side-Channel Analysis for Cryptosystems Based on Supervised Contrastive Learning

Abstract

Deep learning-based profiled side-channel analysis (SCA) targeting cryptographic implementations has attracted significant attention in recent years. Generalizable deep learning mechanisms, such as contrastive learning-based profiled SCA (CL-SCA), can enhance the effectiveness of SCA without reliance on specific network architectures and hyperparameters. This independence enables robust adaptation across diverse attack scenarios. Nonetheless, CL-SCA relies heavily on data augmentation and may mistakenly push apart physical leakage traces that should belong to the same class, which interferes with the extraction of discriminative features crucial for SCA performance. To address these limitations, we propose a profiled SCA method based on supervised contrastive learning, called SupCL-SCA. This method enhances the learning of discriminative features that facilitate key recovery by leveraging supervised information to guide the extraction of similarities in feature space. Compared with state-of-the-art methods, SupCL-SCA not only retains their general applicability and inherent advantages but also eliminates reliance on complex data augmentation and multi-stage training. Additionally, we propose a cosine distance-based Intra-Inter Distance Ratio (IIDR) metric to assess the discriminative capability of models in deep learning-based profiled SCA methods. We evaluate SupCL-SCA on three publicly available datasets covering different implementations and platforms. Experimental results show that SupCL-SCA consistently reduces the number of traces required to recover the key compared to the original methods, demonstrating enhanced attack capability.