Abstract: 【Purpose/significance】With the complexity of social networks,current rumors are often composed of text describing events,corresponding pictures or videos,and multimodal rumors are more likely to convey a wrong cognition to users.The existing research on rumor detection only uses the text features of rumor,and fails to fully explore the relationship between rumor and event【. Method/pro⁃cess】Therefore,this paper proposes a rumor detection method based on enhanced adversary network and multimodal fusion.It uses BERT and Text CNN to extract text features,VGG-19 network to extract image features,and then uses attention mechanism to capture multi-modal feature interaction.Finally,it uses enhanced adversary network to mine the relationship between rumors and events【. Re⁃
sult/conclusion】The experimental results show that the detection accuracy of this method reaches 92.5%, which is about 10~20% higher than that of traditional single-mode and existing multi-mode models【. Innovation/limitation】At present,only the combination of text and image modes has been tried.There are still many short video and voice data on social media.How to integrate more modal fea⁃tures needs to be studied in the future.