Abstract: In complex channel environments， automatic modulation classification in wireless communication places high demand on model performance and reliability. Although deep learning methods have achieved significant progress in this task， their decision-making processes remain “black boxes，” limiting their applicability in critical scenarios. Moreover， existing automatic modulation classification （AMC） model optimization largely relies on human expertise or “black-box” exploration， lacking reliable evidence to guide model decisions and enhancements. To address this， this paper proposes a trustworthy optimization method for integrating AMC attribution analysis and an attention mechanism. Specifically， feature attributions were first obtained using integrated gradients and DeepLIFT， quantifying the contributions of input signal amplitude/phase and in-phase/quadrature channel features to the model output. This allows identification of feature regions that are decisive for classification results. The credibility of these attributions is verified against the physical characteristics of modulated signals. Subsequently， trustworthy attributed features are used as attention weights to optimize the original signals and enhance the model’s responsiveness to decisive feature regions， thereby simultaneously improving classification performance and interpretability. This optimization process relies on trustworthy attributions to ensure consistency between physical signal characteristics and model decisions， thereby enabling credible model optimization. The designed optimization module does not alter the original network structure， ensuring method generality and engineering feasibility. Finally， a signal constellation-based interpretability metric is proposed， to quantify the model’s interpretability through an alignment analysis between attribution and physical features. The experimental results demonstrate that， without changing the network structure， this method improves the classification accuracy of convolutional neural network and long short-term memory models by approximately 12% and 7%， respectively， while achieving a quantifiable correspondence between key features and model decisions.