Optimization of Neural Differential Distinguishers for Simon32/64
俊龙 潘 · Modeling and Simulation · 2025
Simon32/64是美国国安局推荐的轻量级分组密码，现有基于深度学习的差分分析研究多采用单一数据格式和网络模型，未充分挖掘其优化潜力。本文探讨五种输入数据格式与三种神经网络模型的协同效应对Simon32/64神经差分区分器性能的影响。首先，提出倒数第二轮的多三面体输出差分数据格式M3PODPR，并结合四种已有数据格式，与ResNet、带Inception模块的ResNet、SENet进行全组合实验，构建15种数据–模型组合架构，并分别在相同训练集大小和相同明文数据量下，构造神经区分器并进行性能测试。实验表明：在SENet和M3PODPR架构下，9到11轮Simon32/64神经差分区分器均取得最高准确率，优于所有其它组合架构，也优于现有Simon32/64神经区分器的其它结果。因此M3PODPR可有效提升神经区分器的准确率，为密码分析提供新的优化方向。Simon32/64 is a lightweight block cipher recommended by the National Security Agency (NSA). Existing studies on it against deep learning-based differential cryptanalysis primarily adopt a single data format and a network model, failing to fully exploit optimization potential. This paper investigates the synergistic effects of five input data formats and three neural network models on the performance of neural differential distinguishers for Simon32/64. Specifically, we propose the M3PODPR (Multiple 3-Polytope Output Difference Data Format in the Penultimate Round), evaluating the performance of M3PODPR and four existing data formats with ResNet, ResNet with Inception modules and SENet. A total of 15 data-model combinations are constructed and tested under identical training set sizes and plaintexts. Experimental results show that the SENet with M3PODPR architecture achieves the highest accuracy for 9 to 11 rounds of neural differential distinguishers for Simon32/64, outperforming all other combinations and existing results. Therefore, M3PODPR effectively enhances the accuracy of neural distinguishers, providing a new optimization direction for cryptanalysis.