This study aims to verify the effectiveness of the dynamic feature decoupling model in the analysis of cross-era stylistic evolution and its adversarial robustness. In terms of methodology, this paper uses a temporal coupled neural network to construct a joint representation space of style and era, combines adversarial sample generation technology to test the model, and compares the model with the effect of traditional linear correlation analysis. The results show that the accuracy of the model in the task of dating Tang poetry and Song lyrics is improved by 7.6%, and the F1 value of style classification reaches 0.89. However, when faced with dialect interference and semantic attacks, its errors increase by 11.4% and 15.7% respectively. The conclusion shows that the model provides an effective quantitative analysis tool for digital humanities research, but the semantic-level robustness needs to be further improved through adversarial training and corpus expansion.