Paper 2022/1467
A Cipher-Agnostic Neural Training Pipeline with Automated Finding of Good Input Differences
Abstract
Neural cryptanalysis is the study of cryptographic primitives through machine learning techniques. We review recent results in neural cryptanalysis, and identify the obstacles to its application to new, different primitives. As a response, we provide a generic tool for neural cryptanalysis, composed of two parts. The first part is an evolutionary algorithm for the search of single-key and related-key input differences that works well with neural distinguishers; this algorithm fixes scaling issues with Gohr's initial approach and enables the search for larger ciphers, while removing the dependency on machine learning, to focus on cryptanalytic methods. The second part is DBitNet, a neural distinguisher architecture agnostic to the structure of the cipher. We show that DBitNet outperforms state-of-the-art architectures on a range of instances. Using our tool, we improve on the state-of-the-art neural distinguishers for SPECK64, SPECK128, SIMON64, SIMON128 and GIMLI-PERMUTATION and provide new neural distinguishers for HIGHT, LEA, TEA, XTEA and PRESENT.
Metadata
- Available format(s)
-
PDF
- Category
- Attacks and cryptanalysis
- Publication info
- Preprint.
- Contact author(s)
-
emanuele bellini @ tii ae
david gerault @ tii ae
anna hambitzer @ tii ae
matteo rossi @ polito it - History
- 2022-10-26: approved
- 2022-10-26: received
- See all versions
- Short URL
- https://ia.cr/2022/1467
- License
-
CC BY
BibTeX
@misc{cryptoeprint:2022/1467, author = {Emanuele Bellini and David Gerault and Anna Hambitzer and Matteo Rossi}, title = {A Cipher-Agnostic Neural Training Pipeline with Automated Finding of Good Input Differences}, howpublished = {Cryptology ePrint Archive, Paper 2022/1467}, year = {2022}, note = {\url{https://eprint.iacr.org/2022/1467}}, url = {https://eprint.iacr.org/2022/1467} }