基于深度学习的侧信道攻击技术研究综述OA
Survey on Deep Learning-Based Side-Channel Attack
近年来,基于深度学习的侧信道攻击(DL-SCA)凭借其自动特征提取与复杂模式学习能力,成为密码安全领域的重要研究方向.相较于传统侧信道攻击依赖人工特征与统计分析的局限性,借助深度神经网络构建物理泄露信息与密钥操作的非线性关系模型,能显著提升攻击效率与适应性,在应对噪声干扰、掩码防护及去同步化等挑战时优势明显.本文首先介绍了 DL-SCA 的技术框架,明确一般攻击流程及关键阶段划分;接着,梳理了 DL-SCA 中常用多层感知器、卷积神经网络、递归神经网络、生成对抗网络、Transformer 网络等模型结构及其变体的特点及最新研究进展;之后,从功耗、电磁、缓存等不同侧信道领域探讨了 DL-SCA 的应用研究情况.归纳了当前 DL-SCA 面临的模型泛化、超参数调优、可解释性、鲁棒性以及训练成本等关键挑战,并针对每个挑战总结分析了现有的多种解决方案;最后,从场景化技术突破和通用技术体系构建两个维度分析了 DL-SCA 未来的发展趋势,针对小样本、多模态、后量子密码算法、物联网设备 4 种典型场景和跨场景通用技术研究方面给出了具体的技术路径和研究建议.旨在为DL-SCA 技术的理论创新与工程化应用提供聚焦技术本身的深度参考.
Deep learning-based side-channel attacks(DL-SCA)have emerged as a critical research direction in cryptographic security due to their capabilities in automated feature extraction and com-plex pattern learning.Compared to traditional SCA methods relying on manual feature engineering and statistical analysis,DL-SCA leverages deep neural networks to model nonlinear relationships be-tween physical leakage and secret operations,significantly improving attack efficiency and adaptability,particularly against noise,masking,and desynchronization.This paper presents a technical framework of DL-SCA,clarifying its general workflow and key phases.It systematically reviews mainstream model architectures,including multilayer perceptrons,convolutional neural network,recurrent neural network,generative adversarial network,and Transformer,along with their variants and recent ad-vancements.Applications of DL-SCA across power,electromagnetic,and cache-based side channels are analyzed.The study identifies critical challenges currently confronting DL-SCA,including model generalization,hyperparameter optimization,interpretability,robustness,and training cost efficiency.For each challenge,the study synthesizes and analyzes existing mitigation strategies from the literature.Finally,this study examines the future development trends of DL-SCA from two analytical dimen-sions:scenario-specific technological breakthroughs and the construction of a general-purpose technical framework.The study proposes concrete technical pathways and research recommendations for four paradigmatic application scenarios—small-sample contexts,multimodal systems,post-quantum cryp-tographic algorithms,and IoT devices—as well as cross-scenario general-purpose technologies.This systematic analysis aims to provide in-depth technical references centered on core technological inno-vations and their engineering implementations,thereby facilitating both theoretical advancements and practical applications of DL-SCA technology.
严迎建;杨杰;陈琳;蔡爵嵩;郭朋飞
信息工程大学 密码工程学院,郑州 450004信息工程大学 密码工程学院,郑州 450004||中国人民解放军 93656 部队,北京 101100信息工程大学 密码工程学院,郑州 450004信息工程大学 密码工程学院,郑州 450004信息工程大学 密码工程学院,郑州 450004
信息技术与安全科学
侧信道攻击深度学习密码安全神经网络模型鲁棒性强化学习迁移学习
side channel attackdeep learningcryptographic securityneural network modelrobustnessreinforcement learningtransfer learning
《密码学报(中英文)》 2026 (4)
593-611,19
评论