Abstract:

A method was proposed to construct a deep learning distinguisher model for large state block ciphers with large-block and long-key in view of the problem of high data complexity, time complexity and storage complexity of large state block cipher distinguishers, and the neural distinguishers were constructed for SM4 algorithm.Drawing inspiration from the idea that ciphertext difference could improve the performance of distinguishers, a new input data format for neural distinguisher was designed by using partial difference information between ciphertext pairs as part of the training data.The residual neural network model was used to construct the neural distinguisher.The training dataset for large blocks was preprocessed.Additionally, an improved strategy for model relearning was proposed to address the high specificity and low sensitivity of the constructed distinguisher.Experimental results show that the proposed deep learning model for SM4 can achieve 9 rounds neural distinguisher.The accuracy of 4～9 rounds distinguishers can reach up to 100%, 76.14%, 65.20%, 59.28%, 55.89% and 53.73% respectively.The complexity and accuracy of the constructed differential neural distinguisher are significantly better than those of traditional differential distinguishers, and it is currently the best neural distinguisher for the block cipher SM4 to our knowledge.It also proves that the deep learning method is effective and feasible in the security analysis of block cipher of large block.

Key words: block cipher, deep learning, neural distinguisher, SM4 algorithm, complexity