Abstract:

Deep Learning Side Channel Analysis/Attack（DLSCA）based on deep learning is very effective for decryption in various side channel attack scenarios. But DLSCA still has security evaluation problems. This paper is based on the power analysis of AES symmetric encryption algorithm, and explains the reason why traditional machine learning performance metrics such as Accuracy can not evaluate from information entropy. The key information is defined to find out the relationship between side channel security evaluation and the performance of the DNN model during a training phase. Associated with the key information, a DLSCA security evaluation framework is set up.

Key words: side channel analysis, deep learning, security evaluation, information entropy

摘要：

基于深度学习的建模类旁路密码分析（Deep Learning Side Channel Analysis/Attack，DLSCA）对于各种旁路攻击场景的密码破解效果都十分显著，但是DLSCA仍存有安全评估问题。基于AES对称加密算法的能量分析，通过信息熵角度分析准确率等传统机器学习性能指标无法评估DLSCA深度神经网络（Deep Neural Network，DNN）模型训练程度的原因。定义密钥信息量，分别阐释密钥信息量与旁路安全评估、DNN模型训练阶段性能评估的关系，建立深度学习模型与旁路分析二者的联系，提出以密钥信息量为核心的DLSCA安全评估框架。

关键词: 旁路分析, 深度学习, 安全评估, 信息熵