Paper 2021/1017

Improve Neural Distinguisher for Cryptanalysis

Zezhou Hou, Jiongjiong Ren, and Shaozhen Chen

Abstract

At CRYPTO'19, Gohr built a bridge between deep learning and cryptanalysis. Based on deep neural networks, he trained neural distinguishers of Speck32/64 using a plaintext difference and single ciphertext pair. Compared with purely differential distinguishers, neural distinguishers successfully use features of the ciphertext pairs. Besides, with the help of neural distinguishers, he attacked 11-round Speck32/64 using Bayesian optimization. At EUROCRYPTO'21, Benamira proposed a detailed analysis about the inherent workings of Gohr's distinguishers. Although their work opened a new direction of machine learning aided cryptanalysis, there are still two research gaps that researchers are eager to fill in. (1) How to further improve neural distinguishers? (2) Can we conduct effective key recovery on large-size block ciphers adopting neural distinguishers? In this paper, we propose a new algorithm and model to improve neural distinguishers in terms of accuracy and the number of rounds and present effective neural aided attack on large-size block ciphers. First, we design an algorithm based on SAT to improve neural distinguishers. With the help of SAT/SMT solver, we obtain new effective neural distinguishers of SIMON using the input differences of high-probability differential characteristics. Second, we propose a new neural distinguisher model using multiple output differences. Inspired by Benamira's work and data augmentation in deep learning, we use the output differences to exploit more derived features and train neural distinguishers, by splicing output differences into a matrix as a sample. Based on the new model, we construct neural distinguishers of SIMON and Speck with round and accuracy promotion. Utilizing our neural distinguishers, we can distinguish reduced-round NSA block ciphers from pseudo-random permutation better. Moreover, we perform practical key recovery attacks on different versions of SIMON. For SIMON32/64 and SIMON48/96, we append additional 2-round optimal characteristics searched by SAT/SMT solver to the beginning of our neural distinguishers and attack 13-round SIMON32/64, 14-round SIMON48/96 using Gohr's key recovery frame. For SIMON64/128, it costs too much time in precomputation, especially in wrong key response profile, which is unbearable for most of researchers. However, we show with experiments that the distribution of the wrong key profile is pseudo-periodic. Based on this, we make use of partial wrong key profile to describe the whole wrong key response profile, and then propose a generic key recovery attack scheme which can attack large-size block ciphers. As an application, we perform a key recovery attack on 13-round SIMON64/128 using a 11-round neural distinguisher. All our results are confirmed with experiments (source code available online).