Using Machine Learning Approach to Predict Suitable Swaps for the Evolution of S-Boxes

Cryptographic S-boxes are vectorial Boolean functions with specific properties required in cipher design. Given a random vectorial function, we can improve its cryptographic properties with an evolution process. In this study, this is represented by a simple swap of values in the S-box’s vector of values. Our goal is to explore the use of machine learning to create a predictor for a swap selection during the S-box evolution process. We have trained a predictor based on a neural network to predict good swaps with a probability significantly higher than random guessing. Unfortunately, the predictor seems dependent on the dataset, and does not generalize well. Further experiments show that the predictor can be combined with random affine transformations to be usable in the evolution process. However, even in this case, when the S-box is improved beyond the range of the original training data, the prediction performance degrades to essentially a random choice.