Differential neural cryptanalysis against reduced-round SIMON64/96, CHAM64/128, and HIGHT

Abstract

From the perspective of indistinguishability, an attack on a cryptosystem can be modeled as a training process of efficient distinguishers between ciphertexts and random values, or among ciphertexts. Though the theoretical relationship between cryptanalysis and machine learning has been studied and data-driven cryptanalysis methods have been proposed, the attacks become practically available recently due to the progress of the technologies including the parallel processing hardware and the deep learning algorithms. Gohr proposed differential neural cryptanalysis by making neural classifiers learn differential properties of a reduced-round lightweight block cipher in order to obtain the final round key. However, only one 32-bit block cipher called Speck32/64 had been evaluated. In this paper, we train neural distinguishers against three 64-bit reduced-round lightweight ciphers with (generalized) Feistel network, including SIMON64/96, CHAM64/128, and HIGHT, to evaluate learnability and accuracy of the attacks. Various models of distinguishers under different assumptions have been proposed, and the performance of each distinguisher has been empirically assessed.