IDSGA: A high-performance search model for differential-neural cryptanalysis

Differential-neural cryptanalysis poses critical security threats to internet of things-embedded lightweight block ciphers, outperforming traditional methods but requiring efficient input difference identification. Current solutions are constrained by the opacity of neural networks and the prohibitive computational demands during search processes. Therefore, an input difference search model IDSGA is proposed in this paper. IDSGA model innovatively combines cryptographic theory with deep learning through two key processes. First, feature purification enhances differential characteristics by eliminating redundant patterns while preserving attack-relevant statistical properties through traditional differential probability integration. Second, multi-dimensional distribution mapping enables quantitative dataset evaluation by transforming purified data into interpretable statistical metrics. It enables our model to break through the reliance on neural network evaluation. Moreover, IDSGA model constructs a more optimal search path by integrating a genetic algorithm to deal with the problem of differential search of full inputs with high computational complexity. A quantum variant theoretically extends these capabilities. The experimental results show that the execution time of the IDSGA model is 93% less than that of Gohr. For the CARX cipher, the input differences found by IDSGA are more optimal. This work establishes a new paradigm for differential-neural cryptanalysis by decoupling dataset evaluation from neural network training dependencies.