Network Multimedia Rumor Recognition Based On MultiModal Data

With the development of computer technology,the proportion of multimedia content in online rumors is increasing,and multimedia content is more likely to gain the trust of users than plain text content.Therefore,this phenomenon is increasing the difficulty of rumor recognition.The influence of rumors on society is also growing.If it is not properly handled,it will easily lead to serious social unrest.Therefore,how to effectively identify the authenticity of network multimedia content has become a major problem that needs to be solved urgently.Most of the existing network rumor recognition methods only use user data and text data in the published message.The use of image data often only stays on the surface information of the image.In view of the shortcomings of existing identification methods,this thesis proposes a network multimedia rumor recognition method based on multimodal data.the main work and achievements of this thesis include:(1)Introducing image forensics technology into the network rumor recognition system.First,a likelihood probability map describing the possibility of tampering of a JPEG image is extracted by an image tampering region localization algorithm based on block granularity analysis,and the image is divided into a tamper region and a non-tamper region according to the threshold.Secondly,the BAG map of JPEG images is extracted by BAG extraction algorithm.The statistics of BAG graphs are calculated for different regions and the feature vectors describing the degree of image tampering are formed.Finally,the image feature vectors and traditional user data and text data are obtained.The composed multivariate data input rumor recognition model.(2)A two-layer classification model of network multimedia rumors based on image repetition degree is proposed:topic-level rumor classifier and message-level rumor classifier.When constructing the topic-level rumor classifier,first construct the information network topology map with image repetition degree and extract the sub-event theme by community discovery algorithm,then use the topic feature to train the topic-level rumor classifier;when constructing the message-level rumor classifier,In this thesis,an improved CNN network structure aiming at individual differences of users is proposed.Firstly,multiple messages similar to the message to be identified are found through the network topology map.The features are merged with the CNN network structure and then spliced with the to-be-identified message.Finally,the final message-level rumor classifier can be obtained by using its training classifier.By introducing the topic rumor recognition model and improving the CNN network structure,the interference of abnormal points can be reduced and the stability of the model can be improved.In this thesis,the rumor data in the foreign social media platform Twitter is used as the data set used in the research.The experimental results show that the multi-data-based multimedia rumor recognition model proposed in this thesis can better distinguish the rumor content in Twitter,the accuracy of the 2015 VMU mission data set reached 96.7%,and the accuracy of the 2016 VMU mission data set reached 91.2%.Compared with the three benchmark methods,the model's recognition accuracy is better,and the performance is more stable under different data sets.