$eta$-GNN: A Robust Ensemble Approach Against Graph Structure Perturbation

2025-03-26

AI Summary

Graph Neural Networks (GNNs) suffer from insufficient robustness against structural perturbations and rely heavily on prior assumptions about attack patterns. Method: This paper proposes a perturbation-agnostic robust ensemble framework that eliminates such assumptions. Its core innovation is a dynamically learnable weight β, which jointly adapts the fusion strength between GNN and MLP components and explicitly quantifies the degree of input perturbation—thereby unifying robustness and interpretability. The framework integrates dynamic gating, co-training, and perturbation-invariant learning mechanisms. Results: Evaluated on multiple benchmark datasets, the method significantly improves adversarial accuracy under structural attacks. It is the first to enable quantitative assessment of attack severity while strictly preserving original performance on clean data—i.e., zero accuracy degradation on unperturbed inputs.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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 $eta$-GNN, a model enhancing GNN robustness without sacrificing clean data performance. $eta$-GNN uses a weighted ensemble, combining any GNN with a multi-layer perceptron. A learned dynamic weight, $eta$, modulates the GNN's contribution. This $eta$ not only weights GNN influence but also indicates data perturbation levels, enabling proactive mitigation. Experimental results on diverse datasets show $eta$-GNN's superior adversarial accuracy and attack severity quantification. Crucially, $eta$-GNN avoids perturbation assumptions, preserving clean data structure and performance.

Problem

• Enhancing GNN robustness against graph perturbations
• Combining GNN with MLP via dynamic weighting
• Quantifying perturbation levels without clean data sacrifice

Innovation

• Weighted ensemble combining GNN and MLP
• Dynamic weight β modulates GNN contribution
• β indicates and mitigates data perturbation levels