MobileNet-Based Neural Differential Distinguishers for SPECK, GIFT and KATAN

In 2019, Gohr applied deep learning techniques to differential analysis of block ciphers, pioneering a new direction in differential neural cryptanalysis. Building upon Gohr’s Resnet-based differential-neural distinguisher model, this study further explores the use of various neural network models to reduce model parameter count while maintaining the accuracy of neural differential distinguisher. The work in this paper replaces the ResNet structure with the MobileNet model and incorporates depthwise separable convolution technology, significantly reducing the model’s number of parameters while maintaining high accuracy. Experimental results confirm that, while maintaining low computational costs and reduced parameter counts, the distinguishers presented in this study demonstrate a precision comparable to existing research in the differential analysis of lightweight block ciphers SPECK, GIFT, and KATAN. This provides an effective solution for cryptographic analysis on resource-constrained environments.