基于对抗鲁棒增强的多模态谣言检测方法OA

To address the limitations of existing rumor detection datasets being unimodal and the weak robustness of rumor de-tection models against multimodal attacks,this paper constructed a large-scale multimodal rumor dataset,MuRumor.This dataset contained 19 733 posts,658 557 comments,and 12 365 users,covering text,image,comment,user,and social con-text features.Subsequently,it proposed an adversarial robustness-enhanced multimodal rumor detection(AR-MRD)model.The model introduced text style weakening,comment filtering,and image adversarial training mechanisms,combined with a comment-assisted cross-modal consistency mechanism,achieving efficient fusion of multimodal features and significantly impro-ving the model's resistance to cross-modal adversarial attacks.Experimental results show that AR-MRD achieves an accuracy of 92.4％,surpassing existing detection methods,and maintains the lowest attack success rate under adversarial conditions,indicating its effectiveness in defending against tampering attacks on text,image,and comment modalities.