[Resource Topic] 2022/183: Improving Differential-Neural Cryptanalysis with Inception

Welcome to the resource topic for 2022/183

Improving Differential-Neural Cryptanalysis with Inception

Authors: Liu Zhang, Zilong Wang, Baocang wang, Boyang Wang


In CRYPTO’19, Gohr proposed a new cryptanalysis method by building differential-neural distinguishers with neural networks. Gohr combined a differential-neural distinguisher with a classical differential path and achieved a 12-round (out of 22) key recovery attack on Speck32/64. Chen and Yu improved the accuracy of differential-neural distinguisher considering derived features from multiple-ciphertext pairs. Bao et al. enhanced the classical differential path by generalizing the concept of neutral bits, thus launching key recovery attacks for 13-round Speck32/64 and 16-round (out of 32) Simon32/64. Our focus is on improving the accuracy of the distinguisher and training the distinguisher for more rounds using deep learning methods. To capture more dimensional information, we use multiple parallel convolutional layers with kernels of different sizes placed in front of the Residual Network to train differential-neural distinguisher inspired by the Inception in GoogLeNet. For Speck32/64, we obtain a 9-round differential-neural distinguisher and significantly improve the accuracy of distinguishers on 6-,7-, 8-round. For Simon32/64, we obtain a 12-round differential-neural distinguisher and significantly improve the accuracy of the distinguishers on 9-, 10-, and 11-round. In addition, we use neutral bits to solve the same distribution of data required to successfully launch a key recovery attack when using multiple-ciphertext pairs as the input of the neural network. Under the combined effect of multiple improvements, the time complexity of our 11-, 12-, and 13-round key recovery attacks of Speck32/64 is decreased. Also, the success rate of our 12-round key recovery attack reaches 100% in 98 trials. For Simon32/64, we are able to implement a 17-round key recovery attack using the deep learning method for the first time. Also, we decrease the time complexity of the 16-round key recovery attack.

ePrint: https://eprint.iacr.org/2022/183

See all topics related to this paper.

Feel free to post resources that are related to this paper below.

Example resources include: implementations, explanation materials, talks, slides, links to previous discussions on other websites.

For more information, see the rules for Resource Topics .