Fortifying graph neural networks against adversarial attacks via ensemble learning

Chenyu Zhou, Wei Huang, Xinyuan Miao, Yabin Peng, Xianglong Kong, Yi Cao, Xi Chen

Abstract: Highlights•Proposed the Fortifier, which integrates model-based learning to enhance GNNs’ robustness.•Developed the Schur and EEA models as base learners for graph purification.•Introduced GFU as a meta-learner to fuse robust components from Schur and EEA.•Designed a parameter transfer strategy to extend GFU’s robustness to other GNNs.•Extensive experiments confirm Fortifier’s effectiveness across diverse attacks and graph types.