Paper 2020/1235

Assessing Block Cipher Security using Linear and Nonlinear Machine Learning Models

Ting Rong Lee and Je Sen Teh and Jasy Liew Suet Yan and Norziana Jamil and Jiageng Chen

Abstract

In this paper, we investigate the use of machine learning classifiers to assess block cipher security from the perspective of differential cryptanalysis. The models are trained using the general block cipher features, making them generalizable to an entire class of ciphers. The features include the number of rounds, permutation pattern, and truncated differences whereas security labels are based on the number of differentially active substitution boxes. Prediction accuracy is further optimized by investigating the different ways of representing the cipher features in the dataset. Machine learning experiments involving six classifiers (linear and nonlinear) were performed on a simplified generalized Feistel cipher as a proof-of-concept, achieving a prediction accuracy of up to 95%. When predicting the security of unseen cipher variants, prediction accuracy of up to 77% was obtained. Our findings show that nonlinear classifiers outperform linear classifiers for the prediction task due to the nonlinear nature of block ciphers. In addition, results also indicate the feasibility of using the proposed approach in assessing block cipher security or as machine learning distinguishers

Note: This paper is currently under review for a journal publication.