Paper 2020/953

Performance comparison between deep learning-based and conventional cryptographic distinguishers

Emanuele Bellini and Matteo Rossi


While many similarities between Machine Learning and cryptanalysis tasks exists, so far no major result in cryptanalysis has been reached with the aid of Machine Learning techniques. One exception is the recent work of Gohr, presented at Crypto 2019, where for the first time, conventional cryptanalysis was combined with the use of neural networks to build a more efficient distinguisher and, consequently, a key recovery attack on Speck32/64. On the same line, in this work we propose two Deep Learning (DL) based distinguishers against the Tiny Encryption Algorithm (TEA) and its evolution RAIDEN. Both ciphers have twice block and key size compared to Speck32/64. We show how these two distinguishers outperform a conventional statistical distinguisher, with no prior information on the cipher, and a differential distinguisher based on the differential trails presented by Biryukov and Velichkov at FSE 2014. We also present some variations of the DL-based distinguishers, discuss some of their extra features, and propose some directions for future research.

Available format(s)
Secret-key cryptography
Publication info
Preprint. MINOR revision.
distinguisherneural networksTiny Encryption Algorithmdifferential trailscryptanalysis
Contact author(s)
eemanuele bellini @ gmail com
2020-08-13: revised
2020-08-11: received
See all versions
Short URL
Creative Commons Attribution


      author = {Emanuele Bellini and Matteo Rossi},
      title = {Performance comparison between deep learning-based and conventional cryptographic distinguishers},
      howpublished = {Cryptology ePrint Archive, Paper 2020/953},
      year = {2020},
      note = {\url{}},
      url = {}
Note: In order to protect the privacy of readers, does not use cookies or embedded third party content.