Resisting Graph Adversarial Attack via Cooperative Homophilous Augmentation

Recent studies show that Graph Neural Networks (GNNs) are vulnerable and easily fooled by small perturbations, which has raised considerable concerns for adapting GNNs in various safety-critical applications. In this work, we focus on the emerging but critical attack, namely, Injection Attack (GIA), adversary poisons graph injecting fake nodes instead of modifying existing structures or node attributes. Inspired findings adversarial attacks related to increased heterophily perturbed graphs (the tends connect dissimilar nodes), propose a general defense framework CHAGNN against GIA through cooperative homophilous augmentation data model. Specifically, model generates pseudo-labels unlabeled each round training reduce heterophilous edges with distinct labels. The cleaner is fed back model, producing more informative pseudo-labels. such an iterative manner, robustness then promisingly enhanced. We present theoretical analysis effect provide guarantee proposal’s validity. Experimental results empirically demonstrate effectiveness comparison recent state-of-the-art methods diverse real-world datasets.