Paper 2024/322

Theoretical Explanation and Improvement of Deep Learning-aided Cryptanalysis

Weixi Zheng, Xidian University
Liu Zhang, Xidian University
Zilong Wang, Xidian University
Abstract

At CRYPTO 2019, Gohr demonstrated that differential-neural distinguishers (DNDs) for Speck32/64 can learn more features than classical cryptanalysis's differential distribution tables (DDT). Furthermore, a non-classical key recovery procedure is devised by combining the Upper Confidence Bound (UCB) strategy and the BayesianKeySearch algorithm. Consequently, the time complexity of 11-round key recovery attacks on Speck32/64 is significantly reduced compared with the state-of-the-art results in classical cryptanalysis. This advancement in deep learning-assisted cryptanalysis has opened up new possibilities. However, the specific encryption features exploited by DNDs remain unclear. In this paper, we begin by analyzing the features learned by DND based on the probability distribution of a ciphertext pair. Our analysis reveals that DND not only learns the differential features of the ciphertext pair but also captures the XOR information of the left and right branches of the ciphertext pair. This explains why the performance of DND can outperform DDT in certain cases. For other ciphers, we can also predict whether deep learning methods can achieve superior results to classical methods based on the probability distribution of the ciphertext pair. Next, we modify the input data format and network structure based on the specific features that can be learned to train DND specifically. With these modifications, it is possible to reduce the size of their parameters to only 1/16 of their previous networks while maintaining high precision. Additionally, the training time for the DNDs is significantly reduced. Finally, to improve the efficiency of deep learning-assisted cryptanalysis, we introduce Bayes-UCB to select promising ciphertext structures more efficiently. We also introduce an improved BayesianKeySearch algorithm to retain guessed keys with the highest scores in key guessing. We use both methods to launch 11-round, 12-round, and 13-round key recovery attacks on Speck32/64. The results show that under the same conditions, the success rate of 11-round key recovery attacks has increased from Gohr's 36.1% to 52.8%, the success rate of 12-round key recovery attacks has increased from Gohr's 39% to 50%, and the success rate of 13-round key recovery attacks has increased from Zhang et al.'s 21% to 24%. In addition, the time complexity of these experiments is also significantly reduced.

Metadata
Available format(s)
PDF
Category
Attacks and cryptanalysis
Publication info
Preprint.
Keywords
Probability DistributionNeural NetworkParameter AmountBayes-UCBKey Recovery Attack
Contact author(s)
zhengweixi123 @ qq com
liuzhang @ stu xidian edu cn
zlwang @ xidian edu cn
History
2024-02-26: approved
2024-02-25: received
See all versions
Short URL
https://ia.cr/2024/322
License
Creative Commons Attribution
CC BY

BibTeX

@misc{cryptoeprint:2024/322,
      author = {Weixi Zheng and Liu Zhang and Zilong Wang},
      title = {Theoretical Explanation and Improvement of Deep Learning-aided Cryptanalysis},
      howpublished = {Cryptology ePrint Archive, Paper 2024/322},
      year = {2024},
      note = {\url{https://eprint.iacr.org/2024/322}},
      url = {https://eprint.iacr.org/2024/322}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.