Talos: A More Effective and Efficient Adversarial Defense for GNN Models Based on the Global Homophily of Graphs

TL;DR

Talos tackles adversarial robustness for graph neural networks by shifting defense from local, edge-purification strategies to a global-homophily perspective. It formalizes a Global Homophily Index, $Hom = \sum_{k=0}^{\infty} \alpha^{k} Hom^{(k)} = \langle (I-\alpha A)^{-1}, J \rangle$, and derives an edge-removal criterion that maximizes the resulting change in homophily, \Delta Hom. To keep the approach practical for large graphs, it introduces approximations (replacing $M'$ with $M$), batch edge selection from the lower-triangular part of $MJM$, and a matrix-computation strategy for Jaccard similarity, enabling fast preprocessing that does not alter the GNN training. Empirical results across multiple datasets and GNN backbones show that Talos consistently improves accuracy under Metattack and PGD while demanding substantially less runtime than many baselines, including on PubMed-scale graphs. The method is shown to be universal across GNN architectures, though its effectiveness may diminish on graphs with very weak intrinsic homophily. Overall, Talos offers a theoretically grounded, efficient, and broadly applicable defense against large-scale adversarial perturbations in graphs.

Abstract

Graph neural network (GNN) models play a pivotal role in numerous tasks involving graph-related data analysis. Despite their efficacy, similar to other deep learning models, GNNs are susceptible to adversarial attacks. Even minor perturbations in graph data can induce substantial alterations in model predictions. While existing research has explored various adversarial defense techniques for GNNs, the challenge of defending against adversarial attacks on real-world scale graph data remains largely unresolved. On one hand, methods reliant on graph purification and preprocessing tend to excessively emphasize local graph information, leading to sub-optimal defensive outcomes. On the other hand, approaches rooted in graph structure learning entail significant time overheads, rendering them impractical for large-scale graphs. In this paper, we propose a new defense method named Talos, which enhances the global, rather than local, homophily of graphs as a defense. Experiments show that the proposed approach notably outperforms state-of-the-art defense approaches, while imposing little computational overhead.

Figures (4)

• Figure 1: Indirect Attack
• Figure 2: The augmentation experiment on the GCN-Jaccard model
• Figure 3: Model Performance on Pubmed
• Figure 4: Edge distribution histogram under PGD