Improve The Robustness Of Graph Neural Networks Against Adversarial Attacks

Abstract

Graph neural networks (GNN) have achieved remarkable success in many application graph analysis and modeling.

The secret of the great success achieved by GNN in many applications related to graphs is due to the message passing scheme that it adopts during learning, as it collects neighbor messages for each node in each of its layers during training, which allow model in the final  layer to know the graph completely.

Despite the strength of this principle in the tasks of classifying nodes for the graph, GNN's reliance on the graph structure greatly during the message passing makes it vulnerable to adversarial attacks that negatively affect the  Robustness and stability of  these networks and thus a significant decrease in performance and inaccurate results that result in giving nodes A different class from its real classes.