[RE] Are Your Models Still Fair? Fairness Attacks on Graph Neural Networks via Node Injections

TMLR Paper4259 Authors

Abstract: Graph Neural Networks (GNNs) have become indispensable for learning on graph-structured data, with applications in socially sensitive domains such as recommendation systems and healthcare. However, recent research has revealed that fairness-enhancing GNNs remain vulnerable to adversarial attacks, raising concerns about their real-world robustness. This paper represents a reproducibility study of Luo et al. (2024), which demonstrates that adversarial node injection can effectively compromise fairness while preserving overall predictive accuracy. Our results confirm that such attacks are efficient (requiring minimal perturbations), realistic (exploiting feasible node injections), and deceptive (causing fairness degradation without significant accuracy loss). Along with validating the original findings, we redefine their framework as an evasion attack, showing that the attack remains effective on a clean model. Furthermore, we propose a novel defense strategy and analyze the impact of model depth on the attack. Our results highlight the need for more robust GNN architectures against fairness-targeted adversarial threats.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Sheng_Li3

Submission Number: 4259