A machine learning approach to predicting block cipher security

Existing attempts in applying machine learning to cryptanalysis has seen limited success. This paper introduces an alternative approach in applying machine learning to block cipher cryptanalysis. Rather than trying to extract secret keys, machine learning classifiers are trained to predict a cipher's security margin with respect to the number of active s-boxes. Prediction is based on cipher features such as the number of rounds, permutation pattern, and truncated differences. Experiments are performed on a simplified generalised Feistel structure (GFS) block cipher. Prediction accuracy is optimised by refining how cipher features are represented as training data, and tuning hyperparameters. Results show that the machine learning classifiers are able formulate a relationship between the cipher features and security. When used to predict an unseen cipher (a cipher whose data was not used for training), an accuracy of up to 62% was obtained, depicting the feasibility of the proposed approach.