International Association for Cryptologic Research

International Association
for Cryptologic Research


Paper: Enhancing Differential-Neural Cryptanalysis

Zhenzhen Bao , Tsinghua University, China
Jian Guo , Nanyang Technological University, Singapore
Meicheng Liu , Institute of Information Engineering, Chinese Academy of Sciences, China
Li Ma , Institute of Information Engineering, Chinese Academy of Sciences, China
Yi Tu , Nanyang Technological University, Singapore
Search ePrint
Search Google
Presentation: Slides
Conference: ASIACRYPT 2022
Abstract: In CRYPTO 2019, Gohr shows that well-trained neural networks can perform cryptanalytic distinguishing tasks superior to traditional differential distinguishers. Moreover, applying an unorthodox key guessing strategy, an 11-round key-recovery attack on a modern block cipher Speck32/64 improves upon the published state-of-the-art result. This calls into the next questions. To what extent is the advantage of machine learning (ML) over traditional methods, and whether the advantage generally exists in the cryptanalysis of modern ciphers? To answer the first question, we devised ML-based key-recovery attacks on more extended round-reduced Speck32/64. We achieved an improved 12-round and the first practical 13-round attacks. The essential for the new results is enhancing a classical component in the ML-based attacks, that is, the neutral bits. To answer the second question, we produced various neural distinguishers on round-reduced Simon32/64 and provided comparisons with their pure differential-based counterparts.
Video from ASIACRYPT 2022
  title={Enhancing Differential-Neural Cryptanalysis},
  author={Zhenzhen Bao and Jian Guo and Meicheng Liu and Li Ma and Yi Tu},