$β$-GNN: A Robust Ensemble Approach Against Graph Structure Perturbation
Haci Ismail Aslan, Philipp Wiesner, Ping Xiong, Odej Kao
TL;DR
This work addresses the vulnerability of Graph Neural Networks to adversarial graph perturbations by introducing β-GNN, a modular ensemble that combines any GNN with a multi-layer perceptron using a learnable weight β. The core idea is to minimize a robust objective while allowing β to adapt the relative influence of the GNN and MLP, effectively downweighting the GNN under attack and providing a signal for perturbation severity. Empirical results show state-of-the-art or competitive node classification accuracy under adversarial scenarios on both homophilic and heterophilic graphs, with clear interpretability of β dynamics and favorable linear-time scalability. The approach offers practical and proactive robustness advantages for real-world GNN deployments, especially in security-critical systems, while open-source implementation supports reproducibility.
Abstract
Graph Neural Networks (GNNs) are playing an increasingly important role in the efficient operation and security of computing systems, with applications in workload scheduling, anomaly detection, and resource management. However, their vulnerability to network perturbations poses a significant challenge. We propose $β$-GNN, a model enhancing GNN robustness without sacrificing clean data performance. $β$-GNN uses a weighted ensemble, combining any GNN with a multi-layer perceptron. A learned dynamic weight, $β$, modulates the GNN's contribution. This $β$ not only weights GNN influence but also indicates data perturbation levels, enabling proactive mitigation. Experimental results on diverse datasets show $β$-GNN's superior adversarial accuracy and attack severity quantification. Crucially, $β$-GNN avoids perturbation assumptions, preserving clean data structure and performance.