 This work is on enhancing differential neural cryptanalysis. The differential neural cryptanalysis was proposed by Gore in the seminal work in crypto 2019 in the improved key recovery framework. A classical differential is combined with a neural distinguisher, besides a highly selective key guessing strategy was used. In the framework, the neural distinguisher is built by training neural networks to distinguish two types of ciphertext pairs. Pairs whose corresponding plantaxes have a specific difference, and pairs whose plantaxes are randomly selected. The targeted cipher is the smallest version of the lightweight block cipher SPAC3264, which is an ARX cipher. The neural network structure is a commonly used deep residual network. Three training schemes were used to obtain up to eight rounds neural distinguishers. These neural distinguishers are shown to have superior accuracy than the full differential distribution based distinguishers. In this neural distinguishers under the improved key recovery framework, with comparable data complexity, the time complexity of the attack on 11 rounds was reduced up years late. This goes into the next questions. To what extent is the advantage of differential neural cryptanalysis over traditional method, and whether the advantage generally exists in the cryptanalysis of modern ciphers? To answer the first question, we devised differential neural attacks on more rounds of SPAC3264. We found that the previous differential neural attack is not optimal. By enhancing it, the 12-round attack can be improved, and moreover, a practical key recovery attack on 13 rounds can be achieved. Accordingly, differential neural cryptanalysis has more potential than is initially inhibited. The methods developed to enhance the attack are on its classical components. Specifically, there are applications of generalized neutral bits. Neutral bits can be used to generate a structure of data pairs from a single pair, and they conform or do not conform to the prepanded classical differential together, thus they can effectively boost the VIX signal from the distinguisher. For this, we found the simultaneous neutral bit sites, the conditional simultaneous neutral bit sites, and switching bits for adjoining differentials. These neutral bits are not intrinsically linked to neural network-based cryptanalysis, but are expected to be useful for converting a wider range of VIX distinguishers to competitive key recovery attacks. To answer the second question, we produced the RAS neural distinguishers on round reduced statement 3264, and provided comparisons with their four differential distribution table-based counterparts. The comparison results show that R-round neural distinguishers accepting values of the ciphertext pairs achieve similar but weaker classification accuracies than R-round DDT-based distinguishers. We conjecture that R-round ANDIS that accept values of ciphertext pairs could decrypt one unkept round and try to learn the four R-round output different distributions, but fail to learn more features beyond the distribution of differences. It says you've seen the differential neural cryptanalysis framework with device practical key recovery attacks covering 16 rounds. This indicates that differential neural cryptanalysis should work in general on modern ciphers. However, their advantages might be easier to show on ciphers whose differential-like properties cannot be accurately evaluated using existing tools. Thanks for your attention.