Improving the robustness of graph contrastive learning against adversarial attacks via hierarchical medoid-based contrasting.

Abstract

Graph contrastive learning (GCL) has demonstrated its effectiveness in handling graph-structured data through self-supervised learning. However, its robustness against adversarial attacks remains questionable. While this topic is still under debate, a recent empirical study by Guerranti et al. (2023) revealed that among several popular GCL methods, Deep Graph Infomax (DGI) exhibits relatively stronger robustness. This naturally raises an interesting but unaddressed question: what makes DGI essentially more robust than other GCLs against adversarial structure attack? In this paper, we explore the pivot that determines the robustness of graph contrast learning via theoretical analysis of the underlying contrasting mechanism entailed in DGI. In light of our findings, we devise a simple yet robust-improved GCL architecture that can benefit from FIne-grained contrasting with more REliable references (FIRE-GCL), enabling semantic enhancement at the mesoscale level. We validate the effectiveness of FIRE-GCL on both node classification and link prediction tasks across a wide range of graphs. Moreover, we demonstrate the superiority of FIRE-GCL over the state-of-the-art adversarial robustness-oriented GNN models against poisoning attacks. Our study sheds some light on the crucial role of positive samples in defending graph contrastive learning under attack, and can be readily extended to other node-graph contrastive learning variants.