【摘要】 在CRYPTO 2019上, Gohr首次将深度学习技术应用于分组密码的安全性分析,开辟了深度学习辅助密码分析的研究方向.深度学习辅助密码分析的核心为基于深度神经网络构建的差分–神经区分器.相比于经典差分区分器,差分–神经区分器可以达到更高的准确率,有利于降低密钥恢复攻击的数据复杂度.然而,差分–神经区分器的访问涉及大量浮点运算,引入了额外的计算成本,导致攻击的计算复杂度较高.这一缺点极大地限制了深度学习辅助密码分析的性能.本文提出一种通用的深度学习辅助密码分析增强框架,可以在几乎不损失区分准确率的前提下,将差分–神经区分器转换为存储复杂度可忽略的查找表,克服了差分–神经区分器的缺点,显著地增强深度学习辅助密码分析的能力.在美国NSA设计的Speck, Simon以及ISO/IEC标准LEA三类分组密码上的实验,充分地验证了该框架的有效性和通用性.基于该框架,本文改进了对Speck32/64, Speck96以及Speck128的深度学习辅助分析结果.特别地,本文给出了国内外第1个针对13轮Speck32/64完整验证的实际密钥恢复攻击,同时给出Speck96, Speck128目前最长轮数的实际攻击.这些攻击充分说明了本文框架对于增强深度学习辅助密码分析的应用价值.更多还原

【Abstract】 In CRYPTO 2019, Gohr first applied the deep learning techniques to the security analysis of block ciphers, opening the research direction of deep learning-aided cryptanalysis, the core of which is the differential-neural distinguisher based on deep neural networks. Compared with classical differential distinguishers,differential-neural distinguishers can achieve higher accuracy, which can help reduce the data complexity of a key recovery attack. However, the access of a differential-neural distinguisher involves a large number of floating-point operations, bringing an additional computational burden and resulting in a high computational complexity of the attack. This drawback significantly limits the performance of deep learning-aided cryptanalysis. This paper proposes a general enhancing framework for deep learning-aided cryptanalysis, which can convert a differentialneural distinguisher into a lookup table with negligible memory complexity while sacrificing little accuracy,overcoming the shortcoming of the differential-neural distinguisher and enhancing the power of deep learningaided cryptanalysis. Experiments on Speck and Simon block ciphers designed by the National Security Agency of the USA, as well as the ISO/IEC standard LEA block cipher, have fully verified the effectiveness and generality of this framework. With the help of this framework, in this paper, the deep learning-aided cryptanalysis results on Speck32/64, Speck96 and Speck128 are improved. In particular, this paper presents the first fully verified practical key recovery attack on 13-round Speck32/64, as well as the longest practical attacks on Speck96 and Speck128. These attacks fully demonstrate the application value of this framework for enhancing deep learningaided cryptanalysis.更多还原