Research On Robust Training Methods Of GCN Against Topology Attack

Deep neural networks have achieved great success for speech,text and image data,in natural language processing and computer vision.Encouraged by this,researches on Graph Neural Networks around graph data has also attracted widespread attention,and has been gradually applied in recommendation and advertising,among which Graph Convolutional Networks(GCN)is the most representative.GCN captures the neighborhood information of non-Euclidean space through "passing-aggregating"mechanism to obtain stronger feature extraction capabilities,but on the other hand,it also relies heavily on correct topology and is difficult to promote in actual environment.So how improving the robustness on noisy or vulnerable data is current research hotspot of graph neural networks.GCN robustness is influenced by many complex factors,such as data characteristics,attack modes and training methods.Existing robust models often generate a priori knowledge by observing attack behavior,and then infer defense methods to reduce topology attack's damage.These methods using prior hypothesis can not guarantee stability in different datasets and attack modes.Therefore,this thesis focuses on theoretical modeling and general training method of GCN robustness.The specific research contents are as follows:1)We analyze GCN robust training process from information gain and derive optimization expression,which reveals relationship between topology reconstruction and classifier in GCN robust training process.The self-training mechanism,one of semi-supervised learning theory,is used to unify the two in the same optimization objective to build a general robust training framework.2)Considering the ignored graph structure information and low time efficiency under topology reconstruction of general framework,we propose an efficient robust training method based on general framework combined with the difference characteristics of attack modes.For integral-attack,a topology mask method combined with link prediction is proposed to reduce optional edge space,so as to make full use of original graph information and reduce computational complexity.And for target-attack,partial topology reconstruction with neighbor-detection methods is proposed.Finally,the robust training method proposed in this paper is verified by experiments under integral and target attack respectively.Extensive node classification experiments show that the proposed framework has higher accuracy and more stable robustness performance than current GCN robust training methods under different datasets and attack modes.The research of this paper has great instructive for GCN practical application.