基于图数据增强与多模态特征对齐的虚假新闻检测OA

Although researchers have conducted extensive research on multimodal fake news detection,there are still two major challenges:knowledge illusion in large language model(LLM)and cross modal semantic inconsistency.To address the above issues,a fake news detection model GDAFMA based on graph data augmentation and multimodal feature align-ment is proposed.Firstly,the LLM is used to construct multi-dimensional prompt templates for news document style,source authority,and image description,achieving fine-grained semantic deconstruction.Secondly,semantic information is transformed into graph structures,and cross modal deep inference is achieved through graph knowledge enhancement and graph encoding decoding mechanisms.Finally,a multimodal local and global feature alignment strategy is designed to achieve semantic consistency verification and comprehensive credibility evaluation.The experimental results on three publicly available datasets,GossipCop,Weibo,and PolitiFact,show that the classification accuracy of GDAFMA is as high as 0.868,0.916,and 0.898,respectively,far higher than all baseline methods,proving that it can significantly improve the multimodal fake news detection results.