Paper 2020/913

Differential-ML Distinguisher: Machine Learning based Generic Extension for Differential Cryptanalysis

Tarun Yadav and Manoj Kumar

Abstract

Differential cryptanalysis is an important technique to evaluate the security of block ciphers. There exists several generalisations of differential cryptanalysis and it is also used in combination with other cryptanalysis techniques to improve the attack complexity. In 2019, usefulness of machine learning in differential cryptanalysis is introduced by Gohr to attack the lightweight block cipher SPECK. In this paper, we present a framework to extend the classical differential distinguisher using machine learning (ML) based differential distinguisher. We propose a novel technique to construct differential-ML distinguisher for Feistel, SPN and ARX structure based block ciphers. We demonstrate our technique on lightweight block ciphers SPECK, SIMON & GIFT64 and construct differential-ML distinguishers for these ciphers. Data complexity for 9-round SPECK, 12-round SIMON & 8-round GIFT64 is reduced from 2^31 to 2^21, 2^34 to 2^22 and 2^28 to 2^22 respectively. The 12-round differential-ML distinguisher for SIMON is first distinguisher with data complexity less than 2^32.