A Joint Graph Signal and Laplacian Denoising Network

Graph neural network (GNN) models have presented astonishing achievements in various application fields. However, they are shown to be vulnerable to adversarial attacks on graph structure and unnoticeable perturbations on the graph structure can cause significant performance drops in GNN models. Based on recent studies that reveal a class of GNN models is performing graph signal denoising (GSD), in this paper, we design a novel robust GNN model from a joint graph signal and Laplacian denoising problem (GSLD), named GSLDN. Specifically, GSLDN is built based on a block majorization-minimization algorithm for solving the GSLD problem. Designed in such a principled way, GSLDN is endowed with the power to fight against adversarial attacks on graph structure. Experiment results demonstrate the effectiveness of GSLDN.