2026/07/20 by Xue Yuan, Yuan Xue, Qichun Wang
Computer Science · #Artificial neural network #Block (permutation group theory) #Block cipher #Chaos-based Image/Signal Encryption #Coding theory and cryptography #Construct (python library) #Cryptographic Implementations and Security #Differential (mechanical device) #Differential privacy #Key (lock) #Pipeline (software)
paper · doi:10.1093/comjnl/bxag073
published in The Computer Journal (Oxford University Press)
openalex publication_date 2026/07/20 · openalex created_date 2026/07/21 · openalex updated_date 2026/07/22
Abstract At CRYPTO 2019, Gohr pioneered the integration of differential cryptanalysis with neural networks, demonstrating significant advantages over traditional distinguishers. Subsequently, at Inscrypt 2020, Su et al. (Polytopic attack on round-reduced Simon32/64 using deep learning. In: Wu, Y. and Yung, M. (eds.) Information Security and Cryptology: 16th International Conference, Inscrypt 2020, Guangzhou, China. Cham: Springer, 2020, 3–20) proposed polytopic differential neural distinguishers by leveraging multiple effective input differences. More recently, at FSE 2024, Bellini et al. (A cipher-agnostic neural training pipeline with automated finding of good input differences. IACR Trans Symmetric Cryptol 2023;3:184212) introduced a general-purpose tool for automating the training of single-key differential neural distinguishers for various block ciphers. Inspired by these prior works, we extend automated search techniques to related-key differential neural distinguishers, with a focus on jointly identifying effective input and key differences. To this end, we employ a genetic optimization algorithm to automatically discover suitable differential combinations. To validate the effectiveness of the proposed approach, we apply it to the Simeck and Simon cipher families, successfully obtaining useful differential combinations for all three variants of Simeck and all 10 variants of Simon. Furthermore, inspired by the concept of polytopic neural distinguishers, we adopt a novel data representation that leverages multiple distinct input and key differences to construct positive and negative samples, thereby providing the neural network with richer features. Experimental results show that our approach not only identifies high-quality distinguishers for previously unexplored cipher variants, but also achieves higher accuracy than existing state-of-the-art related-key differential neural distinguishers.