Research And Application Of Key Technologies For Detecting Fake News On The Internet

With the rapid development of the Internet,the carrier of news has gradually changed from newspaper to network,and people are more inclined to use computers or mobile phones to receive news.While the Internet brings convenience to people,it also provides a hotbed for the breeding of fake news.With the development of "We-Media",Internet news content is presented in more and more diversified ways,generally containing various information such as text,pictures and comments.How to use such information effectively to detect fake news plays an important role in maintaining social stability and purifying the network space.Fake news is news that can be proved false and spread deliberately.Traditional fake news detection methods usually focus on a single modal information,and the research on fake news detection based on multi-modal is still in its early stage.At present,most detection methods are modeled from the perspective of news text classification or social network transmission.However,due to the rich content of multimedia news,there is still space for improving detection methods based on single modal.Based on this,this thesis takes the microblog multi-modal dataset as an example,proposes a fake news detection framework that integrates multi-modal information,and displays the modeling process and analysis results on the specific dataset.The main work contents and contributions of this thesis are as follows:A multi-modal fake news detection framework,MFND,is proposed to extract features from text,image and user context respectively.Text feature extraction is based on BERT model,and the extracted text features are fine-tuned through the full connection layer to better represent the semantic meaning of news.The Dense Net pre-training model is used to extract the image content convolution features,and the DCT algorithm is used to extract the image frequency domain features to represent the tampering and repeated compression information of the image.User context statistics feature is based on feature engineering,mining user behavior feature and news statistical characteristic.Finally,the features of different modes are spliced and input into the feedforward neural network for training,and the accuracy and F1 value of the final model reach 96.32% and 95.85%respectively.Attention is put forward to improve the fusion mechanism of multimodal fake news attention-MFND detection framework,on the basis of MFND,from the perspective of mining the correlation and the feature crossing between image and text,introducing the attention mechanism,through the image feature vector(query)and the word embedding(key)to distribute value of word in text,resulting from the weighted average of the fusion of the images and text information feature vector,to join the multimodal feature set,model checking has improved effect.The accuracy rate and F1 value of the final model reached 97.91% and 97.52% respectively,which verified the effectiveness of the framework.At the same time,a Web service for fake news detection was developed based on Attention-MFND.