Paper 2025/1443 Original: eprint.iacr.org/2025/1443 ↗

Generic Partial Decryption as Feature Engineering for Neural Distinguishers

Abstract

In Neural Cryptanalysis, a deep neural network is trained as a cryptographic distinguisher between pairs of ciphertexts $(F(X), F(X \oplus \delta))$, where $F$ is either a random permutation or a block cipher, $\delta$ is a fixed difference. The AutoND framework aims to se neural distinguishers that are treated as a generic tool and discourages cipher-specific optimizations. On the other hand, works such as $[\text{LLS}^+24]$ obtain superior distinguishers by adding dedicated features, such as selected parts of the difference in the previous rounds, to the input of the neural distinguishers. In this paper, we study $\text{Generic Partial Decryption}$ as a feature engineering technique and integrate it within a fully automated pipeline, where we evaluate its effect independently of the number of pairs per sample, with which feature engineering is often combined. We show that this technique matches state-of-the-art dedicated approaches on Simon and Simeck. Additionally, we apply it to Aradi, and present a practical neural-assisted key recovery for 5 rounds, as well as a 7-rounds key recovery with $2^{70}$ time complexity. Additionally, we derive useful information from the neural distinguishers and propose a non-neural version of our 5-round key recovery.