DiffGen: a data-driven framework for generating truncated differentials

Differential cryptanalysis involves searching for high-probability differential trails. Traditionally, this search requires the use of constraint solvers or dedicated algorithms. Data-driven methods that rely on machine learning are typically limited to constructing statistical distinguishers for specific ciphers. In this paper, we develop a data-driven approach to the differential search problem by introducing DiffGen, a fully data-driven truncated differential search framework. DiffGen employs a metaheuristic algorithm with an active S-box prediction machine learning model as its fitness function to identify potentially valid truncated differentials within a given range of active S-boxes. A second machine learning model then validates the identified truncated differentials. We demonstrate the effectiveness of the DiffGen framework on generalized Feistel ciphers as a case study. Our results show that DiffGen can effectively generate valid truncated differentials, particularly when using particle swarm optimization as a metaheuristic and a differential validation model based on a fully connected artificial neural network. We verified that 84% of the truncated differentials generated by DiffGen in this setting correspond to actual differential trails. Our findings highlight, for the first time, the feasibility of applying a data-driven approach to the differential search problem.