Differential Cryptanalysis using Reinforcement Learning on round-reduced Ascon

Recently, the automated differential trail search methods are well knowns. Especially for Ascon, using constraint-solving techniques such as MILP, SAT, and CP, which encode mathematical constraints into programs to find optimal solutions. While artificial intelligence-based cryptanalysis has been applied to other ciphers, such as PRESENT, GIFT, and simplified versions of AES, these algorithms tend to have relatively simple structures compared to Ascon. This study proposed reinforcement learning based methodology aim to generate differential trail that the output bits has no differences in the least significant 128-bit of the state, which functions as the authentication tag in finalization phase. We designed concept of the reinforcement learning task as the selection of differences across the 320-bit state of Ascon is independently assigning each of the 64 S-boxes a value ranging from 0 to 31. The policy model is trained in a round-wise manner, with custom reward functions designed for each round to promote the selection of differential trails that correspond with the required input and output differential characteristics. The trained model searched differential trails for the initial target differential pattern, exhibiting no difference in the least significant 128 bits, with probabilities of 2-17 after 2-round and after 3-round.