Deep Learning-Based Differential Distinguishers for Cryptographic Sequences

This research introduces a new deep learning-based technique that significantly enhances the efficiency and accuracy of deep learning based cryptographic distinguishers. By employing a sequence detection approach, we have achieved significant improvements in finding distinguishers for analyzing ciphertext sequences. Two innovative models, an LSTM-based Encoder Classifier (LbEC) and a Transformer based Encoder-only Classifier (TbEC), are proposed. The dataset has been transformed into a list of vector embeddings of the individual sequence data, which is used to train the models. Experimental results demonstrate that this approach has not only achieved results comparable to the existing related works but also outperformed some of the existing schemes. Thereby, distinguishers for HIGHT covering 16 rounds, PRESENT covering 12 rounds, LEA covering 13 rounds, SPARX covering 6 rounds and Piccolo-80 covering 9 rounds have been accomplished, which shows notable improvement over the existing best results.