<![CDATA[Differential analysis is a pivotal method for assessing the security of block ciphers; it distinguishes a cipher from a random permutation by tracing the propagation of plaintext differences. Traditional analytical methods face limitations when applied to complex algorithms, whereas the feature extraction capabilities of deep learning have opened up new avenues for cryptanalysis. To facilitate the security assessment of block ciphers, this paper proposes a novel construction method for a neural differential distinguisher that integrates traditional differential analysis with deep learning techniques. Regarding dataset construction, a multi-ciphertext-pair triplet input format is adopted to preserve differential features while capturing correlations across ciphertext pairs. The network architecture is based on Convolutional Neural Networks (CNNs) and incorporates a Residual Shrinkage Network to construct a deep dilated structure and a multi-scale feature fusion mechanism. Experimental results on the GIFT-128 and ASCON-PERMUTATION lightweight permutation-based cryptographic algorithm demonstrate the efficacy of this approach: for GIFT-128, the 6-round distinguisher reached a maximum accuracy of 99.70%, and the 7-round distinguisher reached 95.47% when using 32 ciphertext pairs; for the 4-round analysis of ASCON, the accuracy rate reached a maximum of 53.54%. These results validate the effectiveness of deep learning methods in the analysis of cryptographic security.]]>
