L2M-GCN: A new framework for learning robust GCN against structural attacks

Abstract

Graph Convolutional Networks (GCNs) have gained extensive attention due to their strong ability to learn from graphs. However, with the advent of stealthy attacks that cause significant differences in node embeddings, the vulnerability of GCNs to malicious attacks has been exposed. Although there are many studies on defense in the spatial or spectral domains, they neglect the complementary roles of the two. In this paper, we propose a new framework, Low frequency and 2-hop in Multi-channel GCN (L2M-GCN), which combines spatial and spectral defense. L2M-GCN has two GCN-based modules. In module one, a new structure reconstructed from learnable spectrum and low-frequency components replaces the adjacency matrix in GCN. In module two, purified 2-hop is introduced and the attention mechanism is used to learn the importance weights of node embeddings. The two modules are eventually assembled into L2M-GCN for joint learning in a parameter-sharing and end-to-end fashion. Extensive experiments demonstrate that L2M-GCN significantly improves the defense performance against structural attacks and outperforms the baselines and state-of-the-art methods.