You are looking at a specific version 20210309:135058 of this paper. See the latest version.

Paper 2021/310

A New Neural Distinguisher Model Considering Derived Features from Multiple Ciphertext Pairs

Yi Chen and Hongbo Yu

Abstract

Gohr has proposed the only deep learning-based distinguisher model at Crypto 2019, which is used to distinguish reduced Speck32/64 and a pseudorandom permutation. This distinguisher model can be applied to many symmetric ciphers. Given a plaintext differential, Gohr’s distinguisher model can learn differences between two distributions from adequate single ciphertext pairs. In this paper, we propose a new neural distinguisher model which takes k > 2 ciphertext pairs as the analysis object. A non-uniform distribution can produce many derived features that will not appear in a single ciphertext pair. Our neural distinguisher model can exploit these derived features from k ciphertext pairs. Taking Gohr’s distinguisher model as the baseline model, we firstly construct strong baseline distinguishers for five reduced ciphers. Then our neural distinguishers for five ciphers are also constructed using the new distinguisher model proposed in this paper. Experiments show our neural distinguishers can always obtain distinguishing accuracy promotions under various settings of k. When combining k samples incorrectly classified by baseline distinguishers into one group, our neural distinguishers can still distinguish correctly with a non-negligible probability. It indicates that derived features have been successfully captured by our neural distinguishers. The distinguishing accuracy promotion also comes from derived features. Our neural distinguishers can also be used to improve the key recovery attack on 11-round Specck32/64. Besides, compared with the raw attack scheme provided by Gohr, we propose a new key recovery attack scheme that can further reduce the time complexity.

Metadata
Available format(s)
PDF
Category
Secret-key cryptography
Publication info
Preprint. MINOR revision.
Keywords
CryptanalysisNeural distinguisherDifferential cryptanalysisDeep learningData reuse
Contact author(s)
chenyi19 @ mails tsinghua edu cn
History
2022-02-24: last of 2 revisions
2021-03-09: received
See all versions
Short URL
https://ia.cr/2021/310
License
Creative Commons Attribution
CC BY
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.