Unraveling Privacy Risks of Individual Fairness in Graph Neural Networks

TL;DR

This paper investigates how improving individual node fairness in Graph Neural Networks can raise privacy risks for edges, revealing a fundamental fairness-privacy trade-off in graph data. It develops PPFR, a two-phase, model-agnostic approach combining fairness-aware loss reweighting via influence functions with privacy-aware graph perturbations during fine-tuning to achieve fairness with limited performance cost and restricted privacy risk. The authors provide theoretical insights and empirical validation on standard GNN benchmarks, demonstrating that PPFR outperforms naive combinations of fairness and privacy methods by achieving favorable bias and privacy metrics with modest accuracy loss. The work advances trustworthy GNNs by addressing both fairness and privacy in a unified framework and suggests future directions for jointly optimizing multiple trustworthiness dimensions in graph learning.

Abstract

Graph neural networks (GNNs) have gained significant attraction due to their expansive real-world applications. To build trustworthy GNNs, two aspects - fairness and privacy - have emerged as critical considerations. Previous studies have separately examined the fairness and privacy aspects of GNNs, revealing their trade-off with GNN performance. Yet, the interplay between these two aspects remains unexplored. In this paper, we pioneer the exploration of the interaction between the privacy risks of edge leakage and the individual fairness of a GNN. Our theoretical analysis unravels that edge privacy risks unfortunately escalate when the nodes' individual fairness improves. Such an issue hinders the accomplishment of privacy and fairness of GNNs at the same time. To balance fairness and privacy, we carefully introduce fairness-aware loss reweighting based on influence function and privacy-aware graph structure perturbation modules within a fine-tuning mechanism. Experimental results underscore the effectiveness of our approach in achieving GNN fairness with limited performance compromise and controlled privacy risks. This work contributes to the comprehensively developing trustworthy GNNs by simultaneously addressing both fairness and privacy aspects.

Figures (7)

• Figure 1: A GNN-based graph database querying system. Given the datum node in a query, the size of other nodes indicates the node similarity score for the given datum. After the vanilla training (i.e., the grey zone), although all other nodes have been accurately predicted by a GNN from the topological view, complaints are raised as the absence of prediction fairness (e.g., for the triangle-indicated node, its size in the prediction space should be larger than the square-indicated node as that in the graph database). To this end, model developers should train GNNs by considering both accuracy and fairness (i.e., the lower green zone). However, privacy attackers can more easily infer confidential edges if GNN fairness is improved.
• Figure 2: In addition to the existing trade-off between performance and fairness/privacy, the trade-off between fairness and privacy is demonstrated in the context of GNNs.
• Figure 3: The framework of privacy-aware perturbations and fairness-aware re-weighting (PPFR) method. After the vanilla training of a GNN model, the privacy-aware module introduces heterophily edges to generate the perturbed graph structure, and the fairness-aware reweighting module employs the influence function and quadratically constrained linear programming (QCLP) to obtain fairness-aware weighted loss. Then, the perturbed graph structure and fairness-aware weighted loss are involved in fine-tuning to promote fairness with limited performance cost and restricted privacy risks.
• Figure 4: Privacy risks of edges before and after enhancing fairness. In this figure, "vanilla"/"Reg" indicates the GCN trained without/with fairness regularisation. The smaller AUC $\Leftrightarrow$ the better privacy.
• Figure 5: The accuracy cost of different methods for fairness on the GCN (left figure) and GAT (right figure) models. In this figure, the higher $\Delta$Acc ($\%$) $\Leftrightarrow$ the better the performance of GNNs.

Theorems & Definitions (6)

• Definition 1: Individual Fairness of Nodes
• Definition 2: Privacy Risks of Edges
• Lemma 5.1
• proof
• Proposition 5.2
• proof