Preimage Attacks Against Lightweight Scheme Xoodyak Based on Deep Learning

With the rapid progress in the artificial intelligence field, machine learning algorithms have been utilized to construct cryptographic schemes and conduct cryptanalysis. In this paper, we propose deep learning-based preimage attacks against variants of Xoodyak hash mode which is a lightweight scheme submitted to the NIST lightweight cryptography standardization project. Three attack models whose internal permutations are of reduced rounds are derived from the original Xoodyak hash mode. Deep neural networks (DNNs) for attack models of 1-round underlying permutations are trained so that the messages of given hash values can be predicted correctly with the networks. In valid attacks, the DNNs are of a low loss rate and high accuracy. This work is more of a tentative attempt to examine the effectiveness of deep learning algorithms employed in conventional preimage attacks. In conclusion, it shows that deep learning techniques make little difference in preimage attacks against Xoodyak hash mode. Compared to the full 12-round internal permutation, only 1 round is covered in the deep learning-based attacks.