Towards unified frameworks for fair and privacy-preserving graph neural networks.

Abstract

Graph Neural Networks (GNNs) deployed in high-stakes applications pose concerns about fairness and privacy. Current methods typically address these issues separately, ignoring the inherent conflict: enforcing individual fairness via distance-based methods unintentionally clusters nodes with similar sensitive attributes, increasing privacy risks such as attribute inference attacks. To resolve this, we propose FPGNN (Fair and Privacy-preserving Graph Neural Network), a unified framework that achieves ranking-based individual fairness using a differentiable ranking approach, and preserves privacy using adversarial training to remove sensitive information from embeddings. For cases with noisy sensitive attributes protected by Local Differential Privacy (LDP), we further propose PL-FPGNN (Private and Limited sensitive attributes-Fair and Privacy-preserving Graph Neural Network), which robustly handles noise via forward correction. Theoretically, we demonstrate that our ranking-based fairness approach avoids privacy leakage amplification inherent in distance-based methods. Extensive experiments on three benchmark datasets verify that FPGNN and PL-FPGNN achieve a good balance among fairness, privacy protection, and predictive accuracy.