Comparative study of Centrality based Adversarial Attacks on Graph Convolutional Network model for Node classification

Graph Convolution Networks (GCNs) are neural networks that can be used to perform different kinds of analysis and mining activities on graphs. GCNs are neural networks that can be directly applied to graphs, and provide an easy way to do node-level, edge-level, and graph-level prediction tasks. They have many practical uses ranging from link prediction to image recognition, from machine translation to node classification. A lot of research is being conducted to research graph information and it has been found that GCNs are also vulnerable to adversarial attacks i.e. the performance of the algorithm can vary drastically based on whether any change has been made to the graph structurally or otherwise. One of the methods of knowing where to conduct the attack on the graph is by using centrality measures, which denote how important a certain node is. For any use case, different kinds of centralities exist which can be used to find the influential nodes in a graph. These measures all work at varying levels of effectiveness depending on their underlying algorithm and the graph structure or features. The Cora citations graph network and a GCN model that performs Node classification task on it is considered here. Before training and classification, uniform edge level adversarial attacks are conducted using four different centrality measures at different scales and the various accuracies are compared.