Self-supervised Adversarial Purification for Graph Neural Networks
Researchers have developed a novel self-supervised adversarial purification framework for Graph Neural Networks (GNNs). This new method separates the task of robustness from classification by using a dedicated purifier, GPR-GAE, which is a graph auto-encoder trained with a self-supervised strategy. The GPR-GAE utilizes multiple Generalized PageRank filters to capture diverse structural representations, enabling effective purification and robust defense against adversarial attacks on graph data. AI
IMPACT Introduces a new method to enhance the security and reliability of Graph Neural Networks against malicious perturbations.