ProGAP Enhances GNN Defense Against Adversarial Attacks.

This research introduces ProGAP, a novel scheme designed to enhance the robustness of Graph Neural Networks (GNNs) against adversarial attacks. GNNs are widely used for modeling complex relationships in various multimedia tasks, but they are vulnerable to perturbations. Current defense mechanisms often suffer from being domain-restricted, leading to insufficient data diversity for learning robust criteria and high computational costs for training new strategies. ProGAP addresses these limitations by leveraging transferable defense knowledge through vulnerability-aware graph prompt learning. It first pretrains a perturbation-capture edge detector on extensive datasets, which jointly models both topological and semantic information to identify universal adversarial patterns. Subsequently, vulnerability-aware prompts are crafted to inject targeted purification guidance into biased nodes, allowing the pretrained detector to adapt to new graph distributions without requiring extensive parameter updates. Experimental evaluations demonstrate that ProGAP significantly improves defense performance, showing a 1-9% gain over state-of-the-art baselines. Furthermore, it achieves this with a substantial reduction in computational time, up to 2.2 times faster. This approach offers a more efficient and broadly applicable solution for securing GNNs in real-world applications.