Deep Learning-Based Differential Distinguishers for NIST Standard Authenticated Encryption and Permutations

Deep learning-based cryptanalysis is one of the new ideas that has emerged in recent years. By using deep learning-based methodologies, researchers are currently modeling conventional differential cryptanalysis. We use deep learning models, CNN, LSTM, LGBM, DenseNet, and LeNet, to generate deep learning-based differential distinguishers that can reveal weaknesses in the encryption schemes. We focus on National Institute of Standards and Technology (NIST) standard lightweight authenticated encryption (AE), such as TGIF-TBC and LIMDOLEN-128, along with permutation methods like SPARKLE-256, ACE-128, and SPONGENT-160. Our research has led us to find that deep learning techniques can generate differential distinguishers for these cryptographic elements. Specifically, we were able to develop differential distinguishers for the SPONGENT-160 permutation up to 7 rounds, for the SPARKLE-256 permutation up to 3 rounds, for the ACE-128 permutation up to 4 rounds, for the TGIF-TBC AE up to 5 rounds, and for the LIMDOLEN-128 AE up to 14 rounds. Notably, this marks the first instance of a deep learning-based differential classifier for the authenticated encryptions TGIF-TBC, LIMDOLEN-128, as well as the permutations SPARKLE-256, ACE-128, and SPONGENT-160, based on our current understanding. When considering various models, both DenseNet and CNN demonstrate strong performance. However, it is the LightGBM (LGBM) model that truly shines as the optimal choice, primarily attributed to its minimal parameter requirements and rapid response speed.