Reassessing Fairness: A Reproducibility Study of NIFA’s Impact on GNN Models
Abstract: Graph Neural Networks (GNNs) have demonstrated exceptional performance in processing graph-structured data, yet fairness concerns remain a critical challenge due to GNNs amplifying bias and prejudice in training data. The Node Injection-based Fairness Attack (NIFA)
(Luo et al., 2024) was recently proposed as a gray-box method to compromise fairness while maintaining model utility. This study aims to reproduce and validate the claims of NIFA, assessing its impact across multiple datasets and GNN architectures. This reproduction
study confirms that NIFA is an effective gray-box attack that degrades the fairness metrics, statistical parity, and equal odds while having a negligible utility loss. Additionally, NIFA’s ability to outperform other graph utility and fairness attacks is inconclusive. Finally, we extend the original work by evaluating NIFA’s performance under multi-class sensitive attributes and varying levels of homophily. NIFA’s ability to degrade fairness shows promising results in a multi-class sensitive attribute environment. Varying levels of homophily showed minimal utility loss and stable fairness metrics across most configurations, with the exception of heterophilic-homophilic and highly homophilic settings. The codebase used in this study can be found at https://anonymous.4open.science/r/Reassessing-NIFA-B4F5/.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Sheng_Li3
Submission Number: 4291
Loading