Topology-Aware Node Injection Attacks against graph neural networks

Linlin Su, Jinyan Wang, Zeming Gan, De Li

Abstract: Graph neural networks (GNNs) are widely applied in real-life scenarios due to their excellent performance in processing graph data. Meanwhile, GNNs are vulnerable to the node injection attack (NIA). The attacker can significantly reduce the effectiveness of GNNs by injecting only a few malicious nodes into the graph. Existing topology construction approaches for NIA are often either random or computationally expensive, which limits the effectiveness of attack strategies and hinders their applicability to large-scale graphs. In this work, we propose a novel NIA method, named Topology-Aware Node Injection Attack (TANIA), which achieves both high attack effectiveness and scalability. TANIA comprises a two-stage topology construction strategy and an adaptive feature optimization module. Specifically, TANIA first selects nodes with weak information-aware ability as the candidate neighbor set to scale down the topology construction search space of the injected nodes. Then, it establishes connections for the injected nodes following the best wrong class consistency strategy to refine the topology for an effective attack with a limited budget. Based on the reconstructed topology, TANIA adaptively optimizes the features of the injected nodes to enhance the attack ability. The experimental results show that TANIA exhibits outstanding attack performance against 14 GNNs compared with eight representative NIA methods while maintaining scalability on large graph scenarios.