FairGFL: Privacy-Preserving Fairness-Aware Federated Learning with Overlapping Subgraphs

Zihao Zhou, Shusen Yang, Fangyuan Zhao, Xuebin Ren

Introduction

Federated Learning (FL) has emerged as a standard for training models on private, distributed data. When extended to graph-structured data—Graph Federated Learning (GFL)—it is common for clients (e.g., financial institutions or social network nodes) to possess overlapping subgraphs, such as shared users or transactions. While data overlap is often viewed as a benefit for reducing heterogeneity, this research identifies a critical downside: utility disparity.

Clients with high degrees of overlap often see significantly better model performance than those with unique data. This "unfairness" discourages participation and hinders global model generalization. FairGFL is introduced as a framework to quantify and mitigate this imbalance through privacy-preserving estimation and adaptive aggregation.

Overview of Graph Federated Learning (GFL) architecture, including Message Passing Neural Networks (MPNN) and the federated client-server communication model.

The FairGFL Framework

FairGFL addresses the unfairness of overlapping subgraphs via three primary components:

  1. Privacy-Preserving Overlap Estimation: To maintain data sovereignty, clients cannot share raw graphs. FairGFL utilizes feature encoders combined with Local Differential Privacy (LDP). Clients sanitize their mini-batch subgraphs using a randomization mechanism before transmission. The server then aggregates these noisy samples to estimate each client's "overlapping ratio" without accessing raw structural data.
  2. Fairness-Aware Aggregation: Unlike standard FedAvg, which treats all updates equally, FairGFL reweights client contributions. "Advantaged" clients with high overlap ratios receive lower weights, while "disadvantaged" clients with unique structural information are prioritized. This prevents the global model from over-optimizing for redundant data regions.
  3. Composite Loss Optimization: To balance fairness with global accuracy, the framework employs a loss-related regularizer ($\lambda ||\vec{F}(w)||_\infty$). This term minimizes the maximum loss among all clients, effectively "pulling up" the performance of the worst-off participants to ensure equitable utility.

Experimental Evaluation

The framework was evaluated on several benchmark graph datasets, including Cora, CiteSeer, Reddit, and large-scale networks like ogbn-arxiv. These datasets were partitioned to simulate varying degrees of structural overlap among federated clients.

Key Results

  • Superior Fairness: FairGFL consistently achieved lower variance and higher entropy in client losses compared to baselines like FedGCN and q-FedAvg. This indicates a more uniform performance distribution across the federation.
  • Convergence and Accuracy: Despite the constraints of fairness and privacy, FairGFL demonstrated faster convergence and higher test accuracy than representative baseline algorithms.
  • Privacy-Utility Trade-off: Experiments showed that even under strict privacy budgets (low $\epsilon$), the server could accurately estimate overlap ratios to maintain fairness without significant utility loss.
  • Efficiency: The communication overhead for sending sanitized subgraphs was minimal, peaking at approximately 8%.
Impact of overlapping ratios on training loss variance and entropy across different datasets, illustrating the fairness challenges in GFL.

Discussion and Future Work

A core insight of this work is the "Graph Unique" problem: because Graph Neural Networks (GNNs) rely on neighborhood structural information, even minor node overlaps can lead to massive redundancy. FairGFL provides an interpretable weighting strategy tied to physical data properties (overlap ratios) rather than black-box metrics.

Future research directions include addressing Multiple Query Attacks, where adversaries might attempt to average out LDP noise over time, and extending the framework to Dynamic Graphs where overlap patterns evolve. Additionally, integrating data quality metrics could further refine the weighting strategy for clients providing noisy or low-quality data.

Summary updated

BibTeX
@article{zhou2026fairgfl,
  author = {Zihao Zhou and Shusen Yang and Fangyuan Zhao and Xuebin Ren},
  title = {{FairGFL: Privacy-Preserving Fairness-Aware Federated Learning with Overlapping Subgraphs}},
  journal = {IEEE Trans. Parallel Distributed Syst.},
  year = {2026},
  doi = {10.1109/tpds.2025.3649863},
}
Figure