Improved differential neural distinguishers for present and skinny

Abstract Lightweight block ciphers can effectively ensure communication security between devices in the Internet of Things (IoT). Differential cryptanalysis is a classical method for evaluating their security. Recently, there has been a notable rise in the use of deep learning techniques for cryptanalysis. To this end, we construct differential neural distinguishers with higher accuracy and stronger generalization ability for the classic cryptographic algorithms PRESENT and SKINNY. First, we combine convolutional neural networks (CNN) with multilayer perceptrons (MLP). This design significantly enhances the synergistic representation of local and global features by optimizing feature interaction mechanisms. Further, we propose a channel attention mechanism called GA-CAM (Global Average and Max pooling-Channel attention module) to improve the model’s focus on critical channel features and better capture channel fine-grained information. In order to enhance the model’s distinguishing capability, we introduce a refined training strategy that combines the RectifiedAdam optimizer with a cyclic learning rate (CLR). Experimental results indicate that our model achieves higher accuracy and, for the first time, obtains 8- and 9-round distinguishers for the SKINNY64/64 cipher. Importantly, our model also demonstrates stronger generalization ability, achieving better accuracy across various input differences.