[Resource Topic] 2020/1235: Assessing Lightweight Block Cipher Security using Linear and Nonlinear Machine Learning Classifiers

Welcome to the resource topic for 2020/1235

Title:
Assessing Lightweight Block Cipher Security using Linear and Nonlinear Machine Learning Classifiers

Authors: Ting Rong Lee, Je Sen Teh, Norziana Jamil, Jasy Liew Suet Yan, 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. These classifiers were trained using common block cipher features (number of rounds, permutation pattern, truncated input and output differences), making our approach generalizable to an entire class of ciphers. Each data sample represents a truncated differential path, for which the level of security is labelled as secure or insecure by the trained classifier based on the number of differentially active S-boxes. We trained six machine learning classifiers (linear and nonlinear) to perform the security prediction task using a dataset generated from a small-scale generalized Feistel structure (GFS) cipher as a proof-of-concept. Prediction accuracy was further refined by determining the best way to represent features in the dataset during training. We then studied how well these classifiers perform the prediction tasks on ciphers that they were trained on (seen) and those that they were not (unseen). When applied on seen ciphers, the classifiers achieved prediction accuracy results of up to 93% whereas for unseen cipher variants, accuracy results of up to 71% were obtained. Our findings indicate that nonlinear classifiers are better suited for the prediction task. Next, we applied the proposed approach to a full-scale lightweight GFS block cipher, TWINE. By training the best performing nonlinear classifiers (k-nearest neighbour and decision tree classifiers) using data from five other GFS ciphers, we obtained an accuracy of up to 74% when labelling data from TWINE. In addition, the trained classifiers could generalize to a larger number of rounds of TWINE despite being trained using data obtained from fewer rounds. These findings showcase the feasibility of using machine learning classifiers, notably nonlinear variants, as a tool to assess block cipher security.

ePrint: https://eprint.iacr.org/2020/1235

See all topics related to this paper.

Feel free to post resources that are related to this paper below.

Example resources include: implementations, explanation materials, talks, slides, links to previous discussions on other websites.

For more information, see the rules for Resource Topics .