GDBA: Defending graph neural networks via attribute debiasing

Zhiqiu Ye, Longlong Lin, Jie Li, Tao Liu, Zeli Wang

Abstract: Highlights•Dual-stage GDBA framework refines graph topology and debiases sensitive attributes.•Attribute-augmented PPR yields up to 39.18% accuracy improvement under attacks.•Integrated gradients identify sensitive features for MI regularization.•Six datasets show consistent robustness gains over state-of-the-art defenses.•Plug-and-play design enables seamless integration into existing GNN models.