Defending Graph Neural Networks via Tensor-Based Robust Graph Aggregation

Jianfu Zhang, Yan Hong, Dawei Cheng, Liqing Zhang, Qibin Zhao

Keywords: Graph Neural Networks, Adversarial Attacks, Robustness, Tensor Decomposition

Abstract: Graph Neural Networks (GNNs) have achieved outstanding success in a wide variety of domains and applications. However, they are still vulnerable to unnoticeable perturbations of graphs specially designed by attackers, causing significant performance drops. Developing algorithms to defend GNNs with robust graphs vaccinating from adversarial attacks still remains a challenging issue. Existing methods treat every edges individually and regularize them by specific robust properties, which ignores the structural relationships among edges and correlations among different properties. In this paper, we propose a tensor-based framework for GNNs to learn robust graphs from adversarial graphs by aggregating predefined robust graphs to enhance the robustness of GNNs via tensor approximation. All the predefined robust graphs are linearly compressed into and recovered from a low-rank space, which aggregates the robust graphs and the structural information in a balanced manner. Extensive experiments on real-world graph datasets show that the proposed framework effectively mitigates the adverse effects of adversarial attacks and outperforms state-of-the-art defense methods.

One-sentence Summary: A tensor-based method to improve robustness of graph neural network models against adversarial attacks.