A Review of Poisoning Attacks on Graph Neural Networks
Kihyun Seol, Yerin Lee, Seungyeop Song, Heejae Park, Laihyuk Park · 2024
A Graph Neural Network (GNN) is designed to generate effective node embeddings in graph-structured data. Therefore, GNNs are well-suited for tasks like node classification and graph generation. As their use has expanded, concerns about their security, robustness, and privacy have grown. This paper explores the various adversarial attacks that can be executed against G NN s, focusing on poisoning attacks, a specific type of adversarial attack. We examine key attack strategies and review recent research developments in each attack. Finally, we conclude with proposals aimed at enhancing the robustness of GNNs.