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

Source: arXiv - 2512.23235v1

Overview

Graph federated learning (GFL) lets multiple parties train a shared graph neural network without exposing their raw graph data. While overlapping subgraphs—nodes and edges that appear in more than one client’s local view—can help mitigate data heterogeneity, the authors show that imbalanced overlaps create fairness problems: clients with fewer shared nodes end up with poorer model performance. The paper introduces FairGFL, a privacy‑preserving, fairness‑aware algorithm that restores balance while keeping the overall predictive quality high.

Key Contributions

• Uncovering unfairness caused by uneven overlapping subgraphs across federated clients, supported by empirical evidence and theoretical analysis.
• Weighted aggregation scheme that uses privacy‑preserving estimates of each client’s overlap ratio to give disadvantaged clients more influence during model merging.
• Fairness‑utility regularizer integrated into the federated loss, explicitly trading off overall accuracy against per‑client fairness.
• Comprehensive evaluation on four real‑world graph benchmarks, demonstrating superior accuracy and fairness compared with four strong baselines.
• Interpretability: the weighting mechanism is transparent, allowing system operators to understand how overlap ratios affect the final model.

Methodology

1. Problem Setting

   • Each client holds a subgraph of a larger global graph. Subgraphs may overlap (share nodes/edges).
   • Overlap ratios differ widely among clients, leading to heterogeneous data quality.

2. Fairness Metric

   • The authors adopt a client‑wise performance disparity measure (e.g., variance of per‑client accuracy) to quantify unfairness.

3. Privacy‑Preserving Overlap Estimation

   • Clients locally compute the size of their overlap with the global graph.
   • Using secure aggregation (e.g., additive secret sharing), the server obtains no‑raw‑data estimates of each client’s overlap ratio.

4. Weighted Model Aggregation

   • Instead of the classic FedAvg (equal weighting), FairGFL assigns higher weights to clients with smaller overlap ratios.
   • The weight (w_i) for client (i) is derived from a monotonic function of its estimated overlap, ensuring the sum of weights equals 1.

5. Fairness‑Utility Regularizer

   • The global loss becomes:
     [ \mathcal{L}_{\text{global}} = \sum_i w_i \mathcal{L}_i + \lambda \cdot \text{FairnessPenalty}({ \mathcal{L}_i }) ]
   • The penalty term penalizes large deviations among client losses; (\lambda) controls the fairness‑utility trade‑off.

6. Training Loop

   • Each round: local GNN training → secure overlap reporting → weighted aggregation with regularizer → updated global model broadcast.

The whole pipeline respects the federated learning privacy guarantees (no raw graph data leaves the client) while adding only modest communication overhead for the overlap statistics.

Results & Findings

• Accuracy boost: FairGFL consistently improves global test accuracy by 2–4 % over vanilla FedAvg and other fairness‑aware baselines.
• Fairness gain: The variance of per‑client performance drops by ~50–70 %, indicating a much more equitable model.
• Ablation studies confirm that both the weighted aggregation and the regularizer are necessary; removing either degrades either fairness or utility.
• Scalability: Communication overhead grows linearly with the number of clients; the extra cost for secure overlap reporting is negligible (< 0.5 % of total traffic).

Practical Implications

• Enterprise Graph Analytics: Companies that jointly train fraud‑detection or recommendation GNNs across siloed data (e.g., banks, e‑commerce platforms) can adopt FairGFL to ensure smaller partners aren’t left with sub‑par models.
• Regulatory Compliance: Fairness‑aware federated learning aligns with emerging AI governance rules that demand equitable outcomes across data contributors.
• Edge‑AI & IoT Networks: In sensor networks where some nodes have limited connectivity (hence fewer overlapping observations), FairGFL’s weighting can compensate without exposing raw sensor readings.
• Open‑Source Tooling: The algorithm can be integrated into existing federated learning frameworks (e.g., Flower, FedML) with minimal changes—just plug in the overlap‑estimation step and replace FedAvg with the weighted aggregator.

Overall, FairGFL provides a ready‑to‑use recipe for developers who need to balance model performance with fairness across heterogeneous graph data owners.

Limitations & Future Work

• Assumption of Honest‑But‑Curious Server: The privacy guarantee hinges on secure aggregation; a malicious server could still infer overlap patterns from repeated weight updates.
• Static Overlap Ratios: The current method treats overlap ratios as fixed per round. In dynamic graphs where edges appear/disappear, the estimation may lag.
• Scalability to Thousands of Clients: Experiments capped at a few dozen clients; further work is needed to validate performance and communication efficiency at massive scale.
• Extension to Heterogeneous Model Architectures: FairGFL assumes all clients train the same GNN architecture. Future research could explore fairness when clients use different model capacities.

The authors suggest exploring differentially private overlap estimation and adaptive weighting schemes that react to temporal changes in graph topology as promising directions.

Authors

• Zihao Zhou
• Shusen Yang
• Fangyuan Zhao
• Xuebin Ren

Paper Information

• arXiv ID: 2512.23235v1
• Categories: cs.LG, cs.DC
• Published: December 29, 2025
• PDF: Download PDF