🤖 AI Summary
Graph Neural Networks (GNNs) exhibit fragile fairness under adversarial perturbations, and existing defenses lack provable robustness guarantees for fairness.
Method: We propose ELEGANT, the first general, plug-and-play framework for provably fair robustness in GNNs. It establishes the theoretical foundation for provable fairness robustness in GNNs, leveraging randomized smoothing coupled with fairness sensitivity analysis—requiring no model modification, retraining, or gradient access, and compatible with arbitrary GNN backbones. Crucially, it achieves fairness certification and bias mitigation jointly without structural assumptions on the graph.
Results: Extensive experiments across multiple real-world datasets and GNN architectures demonstrate that ELEGANT significantly improves fairness robustness against adversarial attacks while simultaneously enhancing the model’s intrinsic bias-mitigation capability.
📝 Abstract
Graph Neural Networks (GNNs) have emerged as a prominent graph learning model in various graph-based tasks over the years. Nevertheless, due to the vulnerabilities of GNNs, it has been empirically proved that malicious attackers could easily corrupt the fairness level of their predictions by adding perturbations to the input graph data. In this paper, we take crucial steps to study a novel problem of certifiable defense on the fairness level of GNNs. Specifically, we propose a principled framework named ELEGANT and present a detailed theoretical certification analysis for the fairness of GNNs. ELEGANT takes any GNNs as its backbone, and the fairness level of such a backbone is theoretically impossible to be corrupted under certain perturbation budgets for attackers. Notably, ELEGANT does not have any assumption over the GNN structure or parameters, and does not require re-training the GNNs to realize certification. Hence it can serve as a plug-and-play framework for any optimized GNNs ready to be deployed. We verify the satisfactory effectiveness of ELEGANT in practice through extensive experiments on real-world datasets across different backbones of GNNs, where ELEGANT is also demonstrated to be beneficial for GNN debiasing. Open-source code can be found at https://github.com/yushundong/ELEGANT.
