Research On Rumor Detection Method Of Social Platform Based On Graph Attention Network And Federated Learning

With the rapid development of the Internet,a lot of mobile data is generated in people’s life.Social platforms not only facilitate people’s daily life,but also generate lots of false rumors which spread along social networks,those rumors cause panic and even harm social stability.In recent years,most researches on rumor detection only propose a rumor detection model for a single social platform,which is often difficult to distinguish the post content of a rumor event with different importance.In addition,rumor detection on a single platform ignores the similarity of rumor features among multiple social platforms,which makes it difficult for multiple platforms to cooperate and jointly conduct rumor detection.Aiming at those research problems existing in appeals,this thesis proposes a rumor detection model based on Graph Neural Network and Federated Learning.The main work contents and innovations are summarized as follows:(1)A rumor detection model based on Bi-Graph Attention Network(Bi-GAT)is proposed to solve the problem of single platform rumor detection.This model first processes the rumor data into graph structure data,and then builds Graph Attention Network models from the two directions of rumor propagation respectively.After the data features of two propagations are aggregated,the classification prediction of rumors will be carried out.This model not only considers the graph structure in the process of rumor propagation,but also aggregates more valuable rumor information through multi-head Graph Attention Network.In addition,in order to prove our model can autonomously focus on valuable information nodes when dealing with rumor detection,this thesis also constructs two kinds of comparative models which can enhance the features of root nodes based on the Bi-GAT rumor detection model.By comparing with other rumor detection models on two public datasets,it is proved that our Bi-GAT rumor detection model has excellent rumor detection performance.(2)A rumor detection model based on Federated Graph Attention Network(Fed GAT)is proposed to solve the problem of multiplatform homogenous rumor data detection.This model combines the horizontal federated learning framework with the Bi-GAT rumor detection model,and takes multiple social platforms as clients in the Fed GAT model.Each social platform construct a Bi-GAT rumor detection model locally,and transmitted the parameter information in the training process to the terminal server in Federated Learning.Then,according to the received aggregation parameters,the local model of each social platform will be fine-tuned.This model can solve the data island problem in the rumor detection of multiple social platforms,and can safely and reliably carry out the rumor detection of multiple platforms.Through simulation experiments,it is verified that the Fed GAT model can yield excellent results when processing datasets from different platforms simultaneously.(3)To solve the problem of heterogeneous rumor data detection on multiple platforms,a heterogeneous rumor data detection model(Per-Fed GAT)based on personalized Fed GAT model is proposed.We propose personalized improvement solutions for heterogeneous rumor data processing from both client and server sides of the Fed GAT rumor detection model.In the client-side model,a correction item representing the difference between the local model and the terminal global model is introduced to reduce the training bias of the local model.In the server-side model,the aggregation of different network layers is improved to make it suitable for heterogeneous rumor data detection.The effectiveness of the Per-Fed GAT model is verified by simulation experiments.The rumor detection model proposed in this thesis can focus on high-value rumor nodes when processing rumor data,and can make full use of Federated Learning to solve the data island problem.The model can be applied to Sina weibo,Wechat and other social network platforms to realize efficient and safe multi-platform rumor detection.In future studies,the application of the rumor detection model proposed in this thesis in a wider range of rumor data types and unbalanced rumor data scenarios can be explored.