Research And Implementation Of Fake News Intelligent Identification Technology Based On Heterogeneous Multi-modal Data

With the rapid development of information technology,more and more people tend to use social networking platforms such as Weibo and Tieba to get news and information.Different from traditional news media,news on social network platforms has the characteristics of low cost of fraud and easy access.Therefore,social platforms have quickly become a major battleground for false public opinion.A large amount of fake information is spread by "malicious users" on Internet social platforms.It poses a threat to Internet security and can even cause social chaos.Due to the low efficiency of artificially identifying fake news,it is of great significance to study and implement an automatic identification and detection method for social fake news.Social media information has the characteristics of large amount of data,diverse structure,and a lot of noise.The use of automated means to identify fake news faces many challenges.Most of the current research work mainly focuses on text content,social context and other types of data,ignoring the important visual data of images,but images often contain rich semantic information,which is conducive to the accurate identification of fake news.Therefore,this paper combines three different structures of data,which are text,image,and social context,and explores the complementary correlations between different modal data,and uses deep learning methods to identify social fake news.The main research contents of this paper are as follows:(1)Research on the feature extraction technology of three modal data of text,image,and social context.At the text feature extraction level,this paper uses the Transformer model of the self-attention mechanism for text representation,which effectively solves the semantic loss of long texts.At the same time,the pre-trained Bert model is used to display and extract the text emotional tendency features;at the image feature extraction level,A multi-scale residual network(Multi-Scaled ResNet,MS-ResNet)based on a convolutional pyramid is proposed for feature extraction of image data.Convolution kernels of different sizes are used to generate feature maps of multiple scales and extract different granularities.The image information can effectively solve the misclassification caused by the loss of features and the incomplete semantic information,and improve the image representation ability;for social context information,this paper proposes the new user fame score FFS feature based on the statistical feature,which is intended to make full use of social Information for auxiliary modeling.(2)Research on multi-modal data fusion methods,and design a multi-modal hybrid fusion strategy based on attention mechanism and Adaboost integration algorithm.The multi-modal hybrid fusion strategy proposed in this paper combines feature-level fusion with decision-level fusion.First,at the feature-level fusion level,text features and image features are modally interacted with through a collaborative attention mechanism,and then with social context features Carry out series splicing to form multi-modal fusion features,and then at the decision-level fusion level,improve the Adaboost algorithm weight update mechanism,introduce the threshold of the number of sample misclassifications to limit the upper limit of the sample weight,and use the improved Adaboost algorithm Perform weighted voting decisions for multiple classifiers.(3)Based on the multi-modal hybrid fusion strategy,a multi-modal fake news identification model Hybrid-MMF(Hybrid Multi-Modal Fake News model,Hybrid-MMF)is proposed.The model first extracts single-modal features from the input multi-modal data,and then uses a multi-modal hybrid fusion strategy to perform heterogeneous fusion of text,image,and social context to obtain the final classification result.This paper uses the MCG-FNeWS'19 data set to conduct experiments.First,a single-modal fake news recognition experiment is performed.The Transformer Encoder text feature extraction network and MS-ResNet image feature extraction network in this paper are respectively compared with related benchmark models.At the same time,Random forest is used to rank the importance of the combined features of social context and text sentiment.The experiment proves the effectiveness of the modal feature extraction method proposed in this paper.Finally,a multi-modal fake news recognition experiment was carried out,and the Hybrid-MMF was compared with predecessor models,multi-modal models with different fusion strategies,etc.The accuracy rate of the model in this paper was 95.7%,which was 2.4%higher than the predecessor model.This article has fully verified the feasibility and advancement of the model in this article from multiple angles by setting up a number of comparative experiments.