Similarity Learning is the most effective way to deal with situations where the relationship between a pair of individuals is a concern. This study primarily investigates the application of Similarity Learning approaches to enhance and accelerate optimization algorithms. In this study, we employed a novel hybridized approach combining the Genetic Algorithm and Similarity Learning to effectively explore and identify optimal S-boxes that exhibit strong nonlinear properties. We have also developed a Siamese Convolutional Neural Network architecture, which is based on a one-dimensional structure consisting of three layers. Furthermore, we have incorporated a novel crossover layer within the Genetic algorithm. A paired dataset was compiled by extracting relevant information from experimental findings. The obtained dataset was further subjected to validation procedures to assess non-linearity. We achieved a nonlinearity score of 110.25 for S-boxes with the initial population for Genetic Algorithm as 10.