A Novel Differential Neural Distinguisher for Cryptanalysis of Lightweight Cipher in IoT

With the growing number of resource-constrained sensors connected to the Internet of Things (IoT), data security during transmission faces significant challenges. Various lightweight ciphers are employed to ensure data integrity. However, key recovery attacks against lightweight ciphers are continually evolving and improving. Recently, neural network-based differential distinguishers have been proposed to enhance the efficiency of key recovery attacks. Nevertheless, the accuracy of these neural network-based distinguishers still needs improvement, especially when targeting cryptosystems with a high number of encryption rounds. In this work, we explore the methods to construct ResNet-based distinguishers with higher accuracy by investigating the residual model, prediction model, and activation functions. First, convolutional models with varying numbers of convolutional layers and different activation functions are designed. Then, a transform layer is introduced to convert the output of the prediction layer into two units. A multi-scale fully connected layer (FCL) is also designed and integrated into the prediction module to improve the accuracy of feature separation. Furthermore, the GELU activation function is used to replace the activation functions in the initial convolutional module and prediction head to achieve higher accuracy. The results demonstrate that the proposed differential neural distinguisher achieves accuracy improvements of approximately 0.5% for single ciphertext pairs and 0.5%–1.83% for multiple ciphertext pairs on Speck 32/64, with even larger improvements observed on other lightweight symmetric ciphers. Additionally, we assess the key recovery attack performance against 11-round Speck 32/64 in an IoT environment. The proposed method improves key-recovery success by 4% with a single ciphertext pair and by 1.9% with multiple pairs. Therefore, the proposed neural distinguisher presents a significant advantage for lightweight ciphers deployed in IoT scenarios.