Integrating Graph Embedding And Privacy Protection Technology For User Relationship Prediction In Social Networks

With the development of technology,social networks have become an important way for people to make friends,interact and exchange information.The social relationships between social users contain important commercial value,and social network relationship prediction can promptly and effectively discover the evolution of social networks,analyze the potential relationships between users,and solve the problem of extensive network structure sparseness.However,it is difficult to handle the social networks with traditional learning algorithms due to their huge dimensions and complex network structure.In addition,social network users are also faced with the risks such as privacy leakage and malicious attacks.Therefore,how to protect the privacy of social network users and achieve efficient relationship prediction has become a research hotspot in the field of social network mining.First of all,in order to solve the problem that traditional learning algorithms are difficult to handle large scale network data,this paper considers the balance theory and state theory in social networks,and designs and implements a graph embedding method SMGAT(Signed Multi-head Graph Attention Network)that combines the multi-head attention mechanism.SMGAT fully considers the signed relationship between user nodes and their first-order and second-order neighbors,and uses a multi-head attention mechanism to give different attention weights to different neighborhood nodes,enriching the neighborhood information of user nodes,and fully mines the potential relationship between users and their neighborhood,improving the performances of relationship prediction.Secondly,in order to effectively prevent social network users from being maliciously attacked,protect the private information of social users,and improve the utility of anonymous data,this paper realizes a k-degree anonymous graph perturbation method BKDA(Balanced K-Degree Anonymity)based on the balance theory,BKDA anonymizes and restructures the graph data through graph modification.In the process of anonymity,BKDA makes social users' node degree values meet k anonymity by structuring degree sequences,for reducing the probability of social users being identified and attacked to 1/k,and protecting users' private information.In the process of graph reconstruction,BKDA combines the balance theory,considers the signed relationships between users and surrounding neighborhoods and retains the original structural characteristics of the network,reduces the changes of the network topology as much as possible,for improving the utility of the anonymous graph data.Finally,in order to effectively integrate differential privacy technology and social network relationship prediction,this paper designs and implements a graph convolutional neural network prediction method DPSPM(Differential Privacy Sign Prediction Method)that incorporates the differential privacy protection mechanism.It makes full use of the graph convolutional network to aggregate and learn the information of the node users and their surrounding neighborhood,improving the utility of the data,and uses the DP-SGD optimizer to add noise that obeys the Gaussian distribution to the loss gradient of the prediction method to achieve differential privacy,preventing attackers from stealing user privacy from the training results,and protecting the privacy of the social users and making effective relationship predictions.