Paper 2024/322

On the Explanation and Enhancement of Neural-inspired Differential Cryptanalysis

Abstract

Neural networks have been applied to symmetric cryptanalysis, and Gohr demonstrated that a neural-network-based distinguisher achieves higher accuracy than classical differential distinguishers at CRYPTO 2019. In this work, we analyze ciphertext data through the lens of probability distributions, identifying non-random features that provide indirect insights into how neural networks distinguish ciphertexts from random data. In parallel, we improve the key-recovery attack process by adopting the Bayesian-UCB method, which achieves a better balance between exploration and exploitation of ciphertext structures. These enhancements reduce the runtime of key-recovery attacks while simultaneously increasing their success rate.