Malicious Account Detection Based On Graph Neural Network In Online Social Networks

Social media has become an indispensable place for people's daily life and social interaction.Billions of users around the world spend much of their day active on social media platforms.These network platforms have become tools for people to communicate and obtain information in real time.However,the openness and convenience of social media also bred many potential dangers.A large number of malicious accounts and false information flooded social networks.Anomaly detection is one of the important data analysis methods to identify normal or abnormal activities on social networks.Graph Convolutional Network(GCN),as a deep learning framework applied to graph data,can be used to detect abnormal accounts and false information on social networks.1)Existing Machine learning-based methods mainly extract the features that can distinguish abnormal accounts from normal accounts by hand.Unfortunately,the training of such models requires a large amount of labeled data,but manual annotation alone is time-consuming and expensive,and the obtained feature vectors are biased and poor in robustness.Graph-based detection methods capture the differences by considering the connectivity among accounts in the social network structure,which cannot converge quickly in large-scale online social network,because they need to simulate a large number of random walks.In response to the above problems,this paper proposes a GCN-based deep autoencoder framework,Capture the complex interaction between each user node and its neighbor nodes through multi-layer graph convolution,Learn the hidden information from the user's high-dimensional attributes and the local structure of the social network,and compress each user node on the social network into a concise low-dimensional embedding vector representation.Further use the learned embedding to reconstruct the original data,enabling us to spot anomalies by measuring the reconstruction error of each node from both the structure and the attribute perspectives.We evaluate our approach against other common approaches on a real-world dataset.Experimental results prove that this method performs well,even best compared to the existing state-of-the-art algorithm.2)Existing rumor detection methods based on machine learning and deep learning usually largely relies on the usage of extensive historical labelled datasets.However,in the early stage of fake news transmission,sufficient forwarding and comment information cannot be obtained.Deep learning-based methods only take into account the patterns of deep propagation but ignore the structures of wide dispersion in fake news detection.In order to tackle this challenge,we propose a weakly supervised fake news detection framework based on graph neural networks.In addition to features related directly to the content of the news,capturing users' profiles and characteristics which can infer the credibility and reliability for each user can also help for fake news detection.Combined content features with user attributes,we apply the inherent aggregation mechanism of the graph neural network to calculate the interaction between a post and its children to realize early detection of fake news.Experiments prove that the GCN model proposed in this paper has significant improvements in false information detection tasks,and is robust in the case of limited forwarded comment data.