Neural-Aided Statistical Attack for Cryptanalysis.

In Crypto'19, Gohr proposed the first deep learning-based key recovery attack on 11-round Speck32/64, which opens the direction of neural-aided cryptanalysis. Until now, neural-aided cryptanalysis still faces two problems: (i) the attack complexity estimations rely purely on practical experiments; (ii) it does not work when there are not enough neutral bits. To the best of our knowledge, we are the first to solve these two problems. In this paper, we propose a Neural-Aided Statistical Attack (NASA) that has the following advantages: (i) NASA supports estimating the theoretical complexity. (ii) NASA does not rely on any special properties including neutral bits. Moreover, we propose three methods for reducing the complexity of NASA. One of the methods, which is based on a newly proposed concept named Informative Bit that reveals an important phenomenon, makes NASA applicable to large-size ciphers. We have performed a series of experiments on round reduced Speck32/64, DES, and Speck96/96. These experiments do not only verify the correctness of NASA, but also further highlight the advantage and potential of NASA. Our work arguably raises a new direction for neural-aided cryptanalysis. [ABSTRACT FROM AUTHOR]