This paper presents improvements to differential-neural cryptanalysis, focusing on enhancing the performance of differential-neural distinguishers and reducing the time and data complexity of key recovery attacks. The authors introduce an Inception module with multiple parallel convolutional layers before the residual block to capture multi-dimensional information, and expand the convolutional kernels in the residual blocks to increase the network's receptive field. These improvements enable the successful training of a 9-round differential-neural distinguisher for SPECK32/64 and a 12-round distinguisher for SIMON32/64. To ensure the required data distribution for key recovery attacks, the authors utilize neutral bits. They also redefine the formula for time complexity based on the differences in prediction speeds between a single-core CPU and a GPU. By combining these advancements, the authors significantly reduce the time and data complexity of key recovery attacks on 13-round SPECK32/64. Additionally, they use knowledge distillation techniques to reduce the model size, accelerating the distinguisher's prediction speed and reducing time complexity. The authors achieve a successful 14-round key recovery attack on SPECK32/64 by exhaustively guessing a 1-round subkey. For SIMON32/64, they accomplish a 17-round key recovery attack for the first time and reduce the time complexity of the 16-round key recovery attack. The paper also discusses the use of multiple classical differentials and SBfADs to further enhance the performance of the differential-neural distinguisher. The results demonstrate the effectiveness of the proposed methods in improving the accuracy and efficiency of differential-neural cryptanalysis.This paper presents improvements to differential-neural cryptanalysis, focusing on enhancing the performance of differential-neural distinguishers and reducing the time and data complexity of key recovery attacks. The authors introduce an Inception module with multiple parallel convolutional layers before the residual block to capture multi-dimensional information, and expand the convolutional kernels in the residual blocks to increase the network's receptive field. These improvements enable the successful training of a 9-round differential-neural distinguisher for SPECK32/64 and a 12-round distinguisher for SIMON32/64. To ensure the required data distribution for key recovery attacks, the authors utilize neutral bits. They also redefine the formula for time complexity based on the differences in prediction speeds between a single-core CPU and a GPU. By combining these advancements, the authors significantly reduce the time and data complexity of key recovery attacks on 13-round SPECK32/64. Additionally, they use knowledge distillation techniques to reduce the model size, accelerating the distinguisher's prediction speed and reducing time complexity. The authors achieve a successful 14-round key recovery attack on SPECK32/64 by exhaustively guessing a 1-round subkey. For SIMON32/64, they accomplish a 17-round key recovery attack for the first time and reduce the time complexity of the 16-round key recovery attack. The paper also discusses the use of multiple classical differentials and SBfADs to further enhance the performance of the differential-neural distinguisher. The results demonstrate the effectiveness of the proposed methods in improving the accuracy and efficiency of differential-neural cryptanalysis.