Resumen en inglés

The search for linear hulls in ARX ciphers can be seen as a sequential decision-making problem: at each step, one selects a transition that affects the final correlation weight. While classical methods solve this through explicit enumeration and pruning, reinforcement learning offers a way to learn a policy that navigates the space of the candidates. This work explores whether such learned guidance can match or improve traditional search baselines. As a basis we use the methods proposed in Mingjiang Huang and Liming Wang's paper Automatic Search for the Linear (Hull) Characteristics of ARX Ciphers: Applied to SPECK, SPARX, Chaskey, and CHAM-64. We investigate the use of reinforcement learning agents trained to find linear hulls for in SPECK32, with the goal of supporting a partial key-recovery attack. For this, instead of using the proposed methods with Matsui's Branch and Bound, we will focus on training a Reinforcement Learning agent on the Combinatorial Linear Approximation table they proposed for SPECK32 by using Q-networks. After this, the agents will be tested with randomly generated masks from a seed. Then they will be compared to the results of the traditional approach to determine which one produces better linear trails to be used in a partial key-recovery attack.