Methodology development for training neural differential distinguishers and key recovery attack on single ciphertext

Key recovery attacks pose a critical threat to cryptographic systems, as these attacks directly compromise the security of encryption mechanisms. The purpose of this study was to develop a novel single-ciphertext key recovery method that addresses the limitations of traditional cryptanalysis techniques in scenarios with limited data availability. The methodology combined an enhanced B-C3-HSwish neural distinguisher with Bayesian optimisation guided by the Upper Confidence Bound (UCB) strategy. The neural distinguisher was improved through multiscale convolutional layers, activation function optimisation, and diversified input structures, while Bayesian optimisation mapped the key recovery problem to a high-dimensional search task using a Gaussian process surrogate model and the UCB acquisition function. Experimental results on the Speck32/64 encryption algorithm demonstrated a 56% success rate in 11 rounds of key recovery attacks, with data complexity ranging from 213.5 to 213.6. This result surpassed the 52.1% success rate obtained in the reference. The B-C3-HSwish neural discriminator has been improved, in particular by replacing the Relu activation function with Hard_swish, to increase its accuracy in singlepair input mode. This study not only offered a new, resource-efficient solution for key recovery in single-ciphertext scenarios, but also highlights the powerful potential of Bayesian optimisation as a promising paradigm for advancing cryptanalysis. The practical value of this work lies in its ability to provide a resource-efficient solution for key recovery in single-ciphertext scenarios, making it particularly suitable for real-world applications with constrained data. This approach also establishes Bayesian optimisation as a promising paradigm for advancing cryptanalysis and enhancing cryptographic security