Privacy Protection Method Of Node Classification Based On Federated Graph Neural Network

In recent years,graph neural networks have made great achievements in graph learning,especially in the field of node classification.Valuable information can be obtained by mining graph data through graph neural networks.Research shows that the existing node classification method based on graph neural network with the method of centralized storage data,however,the reality graph data scattered in different organization,centralized storage data privacy might be the risk,especially privacy will create serious social problems in the field of some sensitive data.Therefore,it is very important to study the privacy protection methods of node classification.Some studies have integrated federated learning into graph neural network and proposed federated graph neural network model,which can effectively protect data privacy.However,in the face of member inference attacks,attackers can infer user privacy information through model parameters.To solve the above problems,this paper adds a differential privacy mechanism to resist member inference attacks,so as to protect data privacy,and on this basis,enrich the client classification label information to improve the performance of the model.The specific research work of this paper is as follows:To solve the problem that federated graph neural network model cannot resist member inference attack,this paper proposes a federated graph neural network model based on local difference privacy.When each client trains the graph neural network,fixed noise is added to the gradient in the back propagation to obtain the disturbed model parameters.The attacker can’t infer the user’s private information from the disturbed model parameters,thus ensuring the security of the private data.However,the accuracy of the model is affected to some extent because adding fixed noise will affect the performance of gradient information.Questions on the above findings,on the basis of the partial differential privacy,put forward a kind of based on adaptive differential privacy federal map neural network model,the client in training neural network diagram,using differential privacy noise minimization algorithm to select the optimal gradient descent step length,minimize the noise impact on gradient,achieved in the protection of privacy at the same time to ensure the accuracy of the models.Because the graph data saved by different clients are not independent and identically distributed in terms of graph structure and node characteristics,the training results of each client model differ greatly,which affects the aggregation effect of global model.To solve the above problems,this paper proposes a federated graph neural network model based on global label,which combines the global label generation method with the federated graph neural network based on differential privacy.The client uploads the disturbed model parameters and node classification prediction results to the server,and the server generates global labels based on the prediction results.The client maps the global labels delivered by the server to the local node to enrich the label information of the node,thus improving the model classification performance.In order to verify the effectiveness of the proposed method,the above three models are trained and tested on the academic citation network data set.Experimental results show that federated graph neural network model based on adaptive differential privacy proposed in this paper has higher accuracy in node classification task than federated graph neural network model based on local difference privacy proposed in local difference privacy.For node classification of non-independent identically distributed data,the model combining global label and differential privacy federated graph neural network proposed in this paper performs better.