A New (Related-Key) Neural Distinguisher Using Two Differences for Differential Cryptanalysis

At CRYPTO 2019, Gohr showed the significant advantages of neural distinguishers over traditional distinguishers in differential cryptanalysis. At fast software encryption (FSE) 2024, Bellini et al. provided a generic tool to automatically train the (related-key) differential neural distinguishers for different block ciphers. In this paper, based on the intrinsic principle of differential cryptanalysis and neural distinguisher, we propose a superior (related-key) differential neural distinguisher that uses the ciphertext pairs generated by two different differences. In addition, we give a framework to automatically train our (related-key) differential neural distinguisher with four steps: difference selection, sample generation, training pipeline, and evaluation scheme. To demonstrate the effectiveness of our approach, we apply it to the block ciphers: Simon, Speck, Simeck, and Hight. Compared to the existing results, our method can provide improved accuracy and even increase the number of rounds that can be analyzed. The source codes are available in https://github.com/differentialdistinguisher/AutoND_New.