[Paper] On the Geometric Coherence of Global Aggregation in Federated GNN

Source: arXiv - 2602.15510v1

Overview

Federated learning lets many devices train a shared model without ever sending raw data to a central server. When the model is a Graph Neural Network (GNN), each client works on its own graph, which can have very different structures and message‑passing dynamics. This paper uncovers a subtle “geometric” failure that occurs when standard federated averaging blindly merges such heterogeneous GNN updates: the global model may look numerically stable (low loss, decent accuracy) but its internal message‑passing behavior becomes incoherent, hurting downstream relational tasks. The authors propose GGRS (Global Geometric Reference Structure), a server‑side regulator that checks the geometric compatibility of client updates before aggregation, preserving the directional consistency of graph transformations without ever looking at private graph data.

Key Contributions

• Identify a new failure mode in cross‑domain federated GNNs: aggregation can destroy the relational geometry of learned message‑passing operators even when conventional metrics stay healthy.
• Formalize “geometric admissibility” for GNN parameters, linking vector‑space representations to the underlying propagation directions, strengths, and sensitivities.
• Introduce GGRS, a lightweight server‑side framework that (1) evaluates the geometric compatibility of each client’s update, (2) rescales or projects updates onto a shared reference subspace, and (3) retains a controlled amount of diversity to avoid over‑regularization.
• Demonstrate empirically on heterogeneous GNN‑native benchmarks and the Amazon Co‑purchase graph that GGRS maintains coherent global message passing across training rounds, improving downstream relational metrics (e.g., link prediction, node classification) without sacrificing privacy.
• Show that standard FL metrics can be misleading for graph‑centric tasks, motivating geometry‑aware monitoring tools for federated graph learning.

Methodology

1. Problem Setup

   • Clients: each holds a private graph (G_i) and trains a local GNN (\theta_i) using its own adjacency and feature matrix.
   • Server: receives model updates (\Delta\theta_i) after a few local epochs and aggregates them (normally via FedAvg).

2. Geometric View of GNN Parameters

   • A GNN layer can be seen as a linear operator that transforms node embeddings based on neighbor aggregation.
   • The direction of this operator determines how information flows across edges, while the magnitude controls sensitivity to neighborhood size.
   • When client graphs differ (e.g., dense vs. sparse, homophilic vs. heterophilic), their learned operators occupy different subspaces in the high‑dimensional parameter space.

3. Geometric Failure Diagnosis

   • The authors compute a geometric similarity score between client updates and a reference structure (e.g., the previous global model’s operator).
   • Low similarity indicates that the client’s update would rotate the global transformation into an incompatible direction, potentially breaking message passing.

4. GGRS Framework

   • Reference Construction: The server maintains a global geometric reference (R_t) (e.g., the average of admissible updates from the previous round).
   • Admissibility Test: For each incoming (\Delta\theta_i), GGRS measures alignment with (R_t) using cosine similarity in the operator‑space.
   • Regulation:
     • If similarity > threshold → accept update unchanged.
     • If similarity < threshold → project (\Delta\theta_i) onto the span of (R_t) (or a learned basis of admissible subspaces) and optionally add a small random component to preserve diversity.
   • Aggregation: The regulated updates are then averaged as usual, producing the next global model (\theta_{t+1}).

5. Privacy Preservation

   • GGRS only needs the parameter vectors; it never accesses raw node features, edge lists, or any client‑side graph topology, keeping the FL privacy guarantees intact.

Results & Findings

• Geometric Coherence: Measured by the average cosine similarity between successive global message‑passing operators, GGRS maintains a stable >0.9 similarity across rounds, whereas FedAvg drifts below 0.6 after a few rounds.
• Training Stability: Loss curves are comparable, but GGRS shows lower variance in validation accuracy, indicating that the model is less prone to “silent degradation.”
• Communication Overhead: GGRS adds <2 % extra bytes per round (a small vector of similarity scores) and negligible compute on the server side.
• Ablation: Removing the projection step (i.e., only threshold‑filtering) yields modest gains, confirming that the geometric projection is the key driver of improvement.

Practical Implications

• More Reliable Federated Graph Services: Companies deploying recommendation engines, fraud detection, or social‑network analytics across edge devices can now trust that the global GNN will retain coherent relational reasoning, even when each device sees a wildly different subgraph.
• Safety‑Critical Applications: In domains like smart grids or autonomous vehicle fleets, inconsistent message passing could lead to erroneous inference about connectivity or risk propagation. GGRS offers a lightweight safeguard without breaking privacy.
• Tooling for FL Engineers: The geometric similarity score can be exposed as a monitoring metric in existing FL dashboards, alerting operators when client updates become too divergent.
• Compatibility: GGRS works as a drop‑in server‑side plug‑in for any FedAvg‑based federated GNN pipeline; no changes are required on the client side.
• Reduced Need for Homogenization: Instead of forcing clients to share graph statistics or to pre‑process data into a common topology, developers can let each client train on its native graph while still achieving a globally consistent model.

Limitations & Future Work

• Assumption of Linear Geometry: The current admissibility test treats GNN layers as linear operators; highly non‑linear architectures (e.g., attention‑based GNNs) may need richer geometric descriptors.
• Threshold Sensitivity: Selecting the similarity threshold requires a small validation sweep; an adaptive or learned threshold could make GGRS more plug‑and‑play.
• Scalability to Massive Client Pools: Experiments used up to 50 clients; future work should evaluate GGRS under thousands of edge devices where the reference structure may become noisy.
• Extension to Heterogeneous Model Architectures: Some federated scenarios allow clients to use different GNN variants (e.g., GCN vs. GraphSAGE). Extending geometric regulation across heterogeneous model families is an open challenge.

The authors suggest exploring curvature‑aware metrics (e.g., Riemannian distances) and integrating GGRS with differential privacy mechanisms as promising directions.

Authors

• Chethana Prasad Kabgere
• Shylaja SS

Paper Information

• arXiv ID: 2602.15510v1
• Categories: cs.LG, cs.DC, cs.NI
• Published: February 17, 2026
• PDF: Download PDF