Deep Learning-Based Rotational-XOR Distinguishers for AND-RX Block Ciphers: Evaluations on Simeck and Simon

The use of deep learning techniques in cryptanalysis has garnered considerable interest following Gohr’s seminal work in 2019. Subsequent studies have focused on training more effective distinguishers and interpreting these models, primarily for differential attacks. In this paper, we shift our attention to deep learning-based distinguishers for rotational XOR (RX) cryptanalysis on AND-RX ciphers, an area that has received comparatively less attention. Our contributions include a detailed analysis of the state-of-the-art deep learning techniques for RX cryptanalysis and their applicability to AND-RX ciphers like Simeck and Simon. Our research proposes a novel approach to identify DL-based RX distinguishers, by adapting the evolutionary algorithm presented in the work of Bellini et al. to determine optimal values for translation ( \(\delta \) ) and rotation offset ( \(\gamma \) ) parameters for RX pairs. We successfully identify distinguishers using deep learning techniques for different versions of Simon and Simeck, finding distinguishers for the classical related-key scenario, as opposed to the weak-key model used in related work. Additionally, our work contributes to the understanding of the diffusion layer’s impact in AND-RX block ciphers against RX cryptanalysis by focusing on determining the optimal rotation parameters using our evolutionary algorithm, thereby providing valuable insights for designing secure block ciphers and enhancing their resistance to RX cryptanalysis.