This study introduces a novel deep learning framework aimed at enhancing the defensive capabilities of Automatic Modulation Recognition (AMR) systems against adversarial attacks through the application of dynamical systems theory and an adaptive weight learning mechanism. Utilizing dynamical systems theory, we analyze and simulate the propagation process of adversarial perturbations, revealing the mechanisms behind various types of attacks, thereby identifying and reinforcing the model's intrinsic resistance to disturbances. At the same time, we introduce an adaptive weight learning mechanism capable of assessing and adjusting the weights of different features, optimizing the feature fusion process to improve the model's accuracy and robustness when facing unknown adversarial samples. Through testing on standard AMR datasets, we have validated that our proposed framework significantly enhances the model's performance under both normal and adversarial conditions. The contributions of this paper lie not only in improving the adversarial defense capabilities of AMR systems but also in providing a new methodology for designing deep learning models using dynamical systems theory and adaptive learning mechanism.