Attention in Differential Cryptanalysis on Lightweight Block Cipher SPECK

The research on combining cryptanalysis with deep learning has recently attracted increasing attention. As an ultra-lightweight cipher for IoT environments, SPECK has attracted much attention from researchers for its excellent performance, and there have been some attempts to introduce deep learning into differential cryptanalysis on SPECK. However, existing work often built differential distinguishers based on traditional residual network, whose accuracy and interpretability on other tasks is inferior to that of attention mechanisms. In order to improve model accuracy and to further utilise the deep learning model to analyse the security of SPECK, this paper introduces the attention mechanism into the differential cryptanalysis on SPECK. First of all, by introducing an attention mechanism in the output layer of the residual network, we achieve a higher accuracy than existing works, and confusion matrices prove that the enhancement brought by attention is effective. Furthermore, using the visualization algorithm, we demonstrate the effectiveness of the attention mechanism intuitively and further analyze the features extracted from the ciphertext by deep learning. In addtion, the bit transfers captured from the ciphertexts by the attention mechanism reflects the possible insecurity of the 5, 6 and 7 rounds of SPECK for specific input differentials, and our work again demonstrates the great potential of deep learning for applications in differential cryptanalysis.