Imperceptible graph injection attack on graph neural networks

Abstract In recent years, Graph Neural Networks (GNNs) have achieved excellent applications in classification or prediction tasks. Recent studies demonstrated that GNNs are vulnerable to adversarial attacks. Modification Attack (GMA) and Injection (GIA) commonly attack strategies. Most graph methods based on GMA, which has a clear drawback: the attacker needs high privileges modify original graph, making it difficult execute practice. GIA can perform attacks without modifying graph. However, many models fail take care of invisibility, i.e., fake nodes be easily distinguished from nodes. To solve above issue, we propose an imperceptible injection attack, named IMGIA. Specifically, IMGIA uses normal distribution sampling mask learning generate node features links respectively, then homophily unnoticeability constraint improve camouflage attack. Our extensive experiments three benchmark datasets demonstrate performs better than existing state-of-the-art methods. As example, shows improvement performance with average increase effectiveness 2%.