Leveraging Synergy to Design Neural Differential Distinguishers for Lightweight Block Ciphers

Researchers have recently become interested in analysing lightweight block ciphers using Artificial Intelligence (AI) techniques. Security against differential cryptanalysis is considered to be a necessary design criterion for block ciphers. In Crypto 2019, Gohr demonstrated that deep learning based neural differential distinguishers (NDDs) for round-reduced versions of SPECK 32/64 yield better accuracy than the traditional differential distinguishers. The present work focuses on building neural differential distinguishers for three families of round-reduced versions of lightweight block ciphers, SPECK (32/64, 64/128), SIMON (32/64, 48/96, 64/128), and PRESENT 64/80, to perform round-reduced neural differential cryptanalysis. We have introduced a novel data fusion technique to leverage possible synergistic interaction between ciphertexts of rounds $(r - 1)$ and round $r$, with an aim to enhance the performance of the NDDs. This fused data, organized in a particular manner, serve as the input for constructing convolutional neural differential distinguishers to differentiate between real pairs and random pairs of ciphertexts. The fusion of the ciphertexts of two successive rounds is done to leverage the merits of not only the ciphertexts of individual rounds but also the synergistic (cooperative) interaction between the ciphertexts of two successive rounds which result in boosting of the performance of the NDDs. The proposed NDDs achieve significantly higher accuracy and can break higher rounds as compared to existing works reported in the literature. Our results demonstrate the presence of synergistic interaction between the ciphertexts of two successive rounds. Specifically, we report results of NDDs on the round-reduced version of SPECK 32/64 upto round 9, SPECK 64/128 upto round 8, SIMON 32/64 with round upto 11, SIMON 48/96 upto round 12, SIMON 64/128 upto round 14 and PRESENT 64/80 upto round 9.