A deep learning-aided key recovery framework for large-state block ciphers
怡 陈, 珍珍 包, 焱天 申, 红波 于 · Scientia Sinica Informationis · 2022
The deep learning-aided key recovery attack is a new cryptanalysis technique published in CRYPTO'2019. The drawback of this technique is that it does not apply to large-state block ciphers. To overcome the drawback, this paper proposes a deep learning-based multistage key recovery framework. The core of this technique is to find a combination of neural distinguishers (NDs) for performing key recovery attacks at each stage. To apply this multistage key recovery framework to large-state members of Speck, multiple NDs are trained and combined into groups. Employing the groups of NDs under the multistage key recovery framework, practical attacks are designed and trialed to show framework effectiveness. The practical attacks are then extended to theoretical attacks, covering more rounds by prepending longer differentials before NDs. Moreover, to boost signals from NDs, an efficient algorithm is proposed to find neutral bits for differentials with low probability. Therefore, considerable improvement is observed in terms of both time and data complexities of differential key recovery attacks on round-reduced Speck with the largest state. This work paves the way for performing cryptanalysis using deep learning on more block ciphers. The related code is available at https://github.com/AI-Lab-Y/NAAF