<p>With the widespread use of social media, while it offers numerous conveniences, it also facilitates the rapid spread of rumors. In recent years, deep learning techniques have made significant strides in identifying the structural features of rumor propagation. However, many rumors now employ various disguise strategies to bypass detection models, rendering most existing approaches unable to accurately identify them. To address these challenges, this paper proposes a more robust rumor detection model from the perspective of graph adversarial training. We introduce Bidirectional Graph Adversarial Training, a rumor detection method built on adversarial training, designed to identify disguised rumors. Our method begins by developing a virtual adversarial feature generation module based on the original propagation structural features to identify the worst-case perturbations. By incorporating adversarial samples during the training process, adversarial training enhances the model’s ability to resist various attacks in real-world scenarios. Specifically, we focus on the influence of node connections to define neighbor perturbations, control the perturbation direction of node features on their neighboring nodes, and introduce adversarial regularizers to defend against the worst-case perturbations. Experimental results on three public datasets demonstrate that our proposed model significantly outperforms state-of-the-art methods, enhancing the robustness of rumor detection.</p>