Robustness And Computational Acceleration Of Graph Neural Networks

The excellent performance of Graph Neural Networks(GNN)on graph-structured data has attracted the interest of researchers.Currently,GNN have various challenges.For example,some studies have shown that GNN,like Convolutional Neural Networks,are very vulnerable to adversarial attacks.In addition,although the message passing mechanism is the powerful cornerstone of GNN,it occupies most of the time and space of GNN reasoning,and is also a hotbed of topological attacks.These problems seriously affect the practical application of GNN.For these problems,we propose corresponding improved algorithms.Our main works are as follows:(1)We find that the Max aggregate function is robust under certain conditions.To satisfy this condition,we design an attention mechanism based on feature similarity according to the characteristics of existing attack methods,thereby proposing a robust message passing mechanism based on feature similarity.In order to further suppress the messages of abnormal nodes,We also propose an adaptive attention mechanism through layer-bylayer filtering and a node feature enhancement algorithm using K nearest neighbors to make the model pay more attention to similar nodes.Based on these improved methods,we propose the Strong Graph Neural Networks(SGNN).Experimental results show that SGNN has good robustness,especially in the case of large perturbations.(2)In order to get rid of the message passing process of GNN in the inference stage,we propose a model named Graph-Less Neural Networks with Neighbouring Contrastive loss(GLNN-NC)based on knowledge distillation,which fully combines the advantages of Graph-Less Neural Networks and Graph-MLP.It can fully learn the representation of nodes while capturing the structural information of the graph,thereby improving the node classification performance.Compared with GNN,GLNN-NC can improve inference speed and fundamentally prevent topological evasion attacks.Extensive experiments verify the effectiveness and advancement of GLNN-NC.