Wang Gao, Wang Gaoli
Shanghai Key Laboratory of Trustworthy Computing, Software Engineering Institute, East China Normal University, Shanghai, North Zhongshan Road, China.
Advanced Cryptography and System Security Key Laboratory of Sichuan Province, Sichuan Province, Chengdu, China.
PeerJ Comput Sci. 2024 Nov 25;10:e2566. doi: 10.7717/peerj-cs.2566. eCollection 2024.
At CRYPTO 2019, Gohr pioneered the application of deep learning to differential cryptanalysis and successfully attacked the 11-round NSA block cipher Speck32/64 with a 7-round and an 8-round single-key differential neural distinguisher. Subsequently, Lu et al. (DOI 10.1093/comjnl/bxac195) presented the improved related-key differential neural distinguishers against the SIMON and SIMECK. Following this work, we provide a framework to construct the enhanced related-key differential neural distinguisher for SIMON and SIMECK. In order to select input differences efficiently, we introduce a method that leverages weighted bias scores to approximate the suitability of various input differences. Building on the principles of the basic related-key differential neural distinguisher, we further propose an improved scheme to construct the enhanced related-key differential neural distinguisher by utilizing two input differences, and obtain superior accuracy than Lu et al. for both SIMON and SIMECK. Specifically, our meticulous selection of input differences yields significant accuracy improvements of 3% and 1.9% for the 12-round and 13-round basic related-key differential neural distinguishers of SIMON32/64. Moreover, our enhanced related-key differential neural distinguishers surpass the basic related-key differential neural distinguishers. For 13-round SIMON32/64, 13-round SIMON48/96, and 14-round SIMON64/128, the accuracy of their related-key differential neural distinguishers increases from 0.545, 0.650, and 0.580 to 0.567, 0.696, and 0.618, respectively. For 15-round SIMECK32/64, 19-round SIMECK48/96, and 22-round SIMECK64/128, the accuracy of their neural distinguishers is improved from 0.547, 0.516, and 0.519 to 0.568, 0.523, and 0.526, respectively.
在2019年密码学大会上，戈尔率先将深度学习应用于差分密码分析，并成功使用一个7轮和一个8轮的单密钥差分神经区分器攻击了11轮的美国国家安全局分组密码Speck32/64。随后，卢等人（DOI 10.1093/comjnl/bxac195）提出了针对SIMON和SIMECK的改进相关密钥差分神经区分器。在此工作之后，我们提供了一个框架来构建针对SIMON和SIMECK的增强型相关密钥差分神经区分器。为了有效地选择输入差分，我们引入了一种利用加权偏差分数来近似各种输入差分适用性的方法。基于基本相关密钥差分神经区分器的原理，我们进一步提出了一种改进方案，通过利用两个输入差分来构建增强型相关密钥差分神经区分器，并且在SIMON和SIMECK上都获得了比卢等人更高的准确率。具体而言，我们对输入差分的精心选择使得SIMON32/64的12轮和13轮基本相关密钥差分神经区分器的准确率显著提高了3%和1.9%。此外，我们的增强型相关密钥差分神经区分器优于基本相关密钥差分神经区分器。对于13轮的SIMON32/64、13轮的SIMON48/96和14轮的SIMON64/128，其相关密钥差分神经区分器的准确率分别从0.545、0.650和0.580提高到了0.567、0.696和0.618。对于15轮的SIMECK32/64、19轮的SIMECK48/96和22轮的SIMECK64/128，其神经区分器的准确率分别从0.547、0.516和0.519提高到了0.568、0.523和0.526。
