Paper 2020/1620

Neural Aided Statistical Attack for Cryptanalysis

Yi Chen and Hongbo Yu

Abstract

Gohr improved attacks on 11-round Speck32/64 using deep learning [17] at Crypto 2019, which is the first work of neural aided cryptanalysis. But we find that the key recovery attack model proposed by Gohr is limited by several properties. It relies heavily on the neutral bit which doesn’t always exist in the attacked cipher. Besides, the data complexity, computation complexity, and success rate can only be estimated through the practical attack. In this paper, we propose a neural aided statistical attack that can be as generic as the differential cryptanalysis. It has no special requirements about the attacked cipher and allows us to estimate the theoretical complexities and success rate. For reducing the key space to be searched, we propose a Bit Sensitivity Test to identify which ciphertext bit is informative. Then specific key bits can be recovered by building neural distinguishers on related ciphertext bits. Applications to round reduced Speck32/64, Speck48/72, Speck48/96, DES prove the correctness and superiorities of our neural aided statistical attack.