FairDP-GNN: Graph Neural Network with Group Fairness and Differential Privacy

Graph Neural Networks (GNNs) have shown excellent performance in learning node representations of various types of graph-structured data. However, GNNs may capture sensitive information in the raw data, leading to privacy leakage of nodes or making unfair decisions for specific groups undermining group fairness. Differential Privacy (DP), as a widely accepted privacy-enhancing technology, is increasingly applied in machine learning to ensure privacy protection. Unfortunately, adding DP to GNNs has negative impacts on group fairness, resulting in discriminatory decisions from them and thus reducing user trust. To address the above problems, we propose a fair GNN training method based on DP called Fair Differential Privacy Graph Neural Network (FairDP-GNN). Specifically, we designed an Aggregation Differential Privacy Mechanism (ADPM) to impose the noise to the aggregation in the GNNs, thereby ensures the effectiveness and reasonableness of the noise. Furthermore, we developed a Group Fairness Promotion Strategy (GFPS) to improve the group fairness by counteracting or exploiting the effects of noise that may exacerbate or mitigate group bias. Experimental results on three real-world datasets show that our method can realize a good balance among model utility, privacy preservation, and fairness.