Neural-Linear Attack Based on Distribution Data and Its Application on DES

Abstract

The neural-differential distinguisher proposed by Gohr boosted the development of neural aided differential attack. As another significant cryptanalysis technique, linear attack has not been developing as rapidly in combination with deep learning technology as differential attack. In 2020, Hou et al. proposed the first neural-linear attack with one bit key recovery on 3, 4 and 5-round DES and restricted multiple bits recovery on 4 rounds, where the effective bits in one plain-ciphertext pair are spliced as one data sample. In this paper, we compare the neural-linear cryptanalysis with neural-differential cryptanalysis and propose a new data preprocessing algorithm depending on their similarities and differences. We call the new data structure distribution data. Basing on it, we mount our key recovery on round-reduced DES—first, we raise the accuracy of the neural-linear distinguisher by about 50%. Second, our distinguisher improves the effectiveness of one bit key recovery against 3, 4 and 5-round DES than the former one, and attack 6-round DES with success rate of 60.6% using 2048 plain-ciphertext pairs. Third, we propose a real multiple bit key recovery algorithm, leading neural-linear attack from theory to practice.