Research On Abnormal User Detection And Privacy Protection In Online Social Networks

The rapid development of social network has brought a lot of traffic and business value,but also attracted a lot of abnormal users who profit from improper behavior.These abnormal users often evade detection by disguising their behavior,but cannot easily manipulate user interactions.Therefore,the interaction information between users can be considered in detecting abnormal users.Real social networks are asymmetric and heterogeneous networks with multiple relationships.However,many current studies only consider single-relationship social networks and simplify the interaction relationships to symmetric ones.In addition,when social network data is used for research work,there is a risk of user privacy information disclosure.Therefore,before releasing the data set to the third party,the data owner should protect the privacy of user information and realize the anonymity of social network data.In order to deal with the disguised behavior of abnormal users,an abnormal user detection mechanism ADMSDG(Abnormal user detection in multi-relationship social Networks based on dynamic weighted graph con-volution)for multi-relationship social networks is designed.In order to improve the scalability of the method,the knowledge graph embedding technology is used in the representation learning layer to automatically learn the semantic information of various interaction relationships in the network;the traditional graph convolutional neural network does not distinguish the importance of neighbor nodes,nor can it capture the difference between the active sending and passive receiving of interactive behaviors,so the iterative results are used as prior knowledge,the asymmetry of the interaction relationship is considered,and the dynamic weighted graph convolution operator is designed in combination with the trust propagation rules.The experimental results show that abnormal user detection can be realized based on various interactive behavior information between users in social networks,and the asymmetric performance of interaction relationships can be considered to effectively improve the performance of the model.The classification of users can be predicted by learning the information of various interaction relationships in social networks,indicating that the representation learning of interaction relationships is effective in learning user tag information.The tag information of users in the iteration results is helpful to further mining the information of interaction relations,which indicates that the tag information of users is also effective in learning the representation of interaction relations.Therefore,considering the intrinsic correlation between semantic information of interaction relationship and user tag information,information sharing between tasks is realized based on multi-task learning framework.A new Abnormal user detection mechanism ADCTR(Abnormal user Detection based on internal correlation between tag information and social relationship)is designed for multirelationship social networks.Experimental results show that some potential information can be effectively learned by considering the internal correlation between interactive relationship information and user label information,which further improves the detection performance of the model.In order to solve the problem of user privacy leakage in the research of abnormal user detection and reduce the information loss in the process of privacy protection,a k-degree anonymous privacy protection method AKDA(Available K-Degree-Anonymity)is designed.The least modified anonymous sequence was found by depth traversal and pruning.In order to reduce structural damage and information loss in the process of reconstruction,the social graph is reconstructed based on the structural information of the original social network and the background knowledge of abnormal user detection.Experimental results show that under the same privacy protection,AKDA method can effectively reduce information loss and improve data availability.