Federated Learning for Graph Structured Data

28 concepts22 papers1 proposal

Standard federated learning (FL) operates on an elegant assumption: data samples are independent and isolated. Whether training on images across smartphones or text documents across hospitals, each client possesses self-contained data points.

Graph data completely breaks this assumption. In a graph, the data points (nodes) are explicitly defined by their relationships to other data points (edges). As surveyed by Zhao et al. and Wu et al., centralizing sensitive graphs—such as financial transactions, medical networks, or social circles—violates modern privacy regulations. Federated Graph Learning (FedGL) addresses this challenge, enabling multiple clients to collaboratively train Graph Neural Networks (GNNs) by exchanging only model updates (weights or gradients) rather than raw adjacency matrices or node features, as demonstrated by Kim et al..

The core technical hurdle of FedGL is the edge boundary problem: how can a GNN perform message passing when a node's neighbors reside on a completely different, private client?

Bridging Broken Edges Across Clients

In a horizontally partitioned setting, different clients hold distinct subgraphs or individual nodes. Standard GNN layers compute a node's next-state representation by aggregating features from its immediate neighborhood:

$$h_v^{(l+1)} = \text{AGGREGATE}\left(\{h_u^{(l)} : u \in \mathcal{N}(v)\}\right)$$

If neighbor $u$ belongs to Client B, Client A cannot compute this update without exposing raw connectivity.

Client A (Private)          Client B (Private)
  [ Node V ]                  [ Node U ]
      \                            /
       \__ Broken Edge Boundary __/

To circumvent this barrier without leaking raw data, Wu et al. developed an anonymous neighbor expansion protocol. By leveraging a secure third party to match encrypted identifiers, clients can securely fetch and incorporate embeddings from structural neighbors on other devices without exposing the underlying user-item interactions to the primary server.

Alternatively, frameworks can eliminate cross-client edge communication entirely by predicting missing structural context locally. Pan et al. designed a framework where clients maintain strict sovereignty over their local connectivity but utilize a differentially private hierarchical clustering tree alongside a 2-hop neighbor predictor. This allows the local model to approximate global structural properties without explicitly querying external nodes. In communication networks, Xiao et al. demonstrated that decentralized GCNs can collaboratively learn a shared reasoning model, allowing nodes to automatically infer hidden relations from limited local "clues" rather than relying on a centralized repository.

Disentangling Global and Local Graph Components

Another key insight behind FedGL is that graph structures can be mathematically factored into shared global properties and private local variations.

For Knowledge Graphs (KGs), which consist of entity-relation-entity triples, Chen et al. introduced a framework that decouples entity representations from relation representations. While entity embeddings are aggregated globally by a central server to capture universal semantic meanings, the relation embeddings and the underlying graph triples remain strictly local to each client. This ensures that the global model learns what entities are, while individual clients protect how those entities interact within their private silos.

Global Server:  [ Aggregated Entity Embeddings ] (Shared Vocabulary)
                       ▲                ▲
                       │                │
Client 1:       [ Local Relations ]   Client 2: [ Local Relations ]
                [ Private Triples ]             [ Private Triples ]

In spatio-temporal domains, such as skeleton-based action recognition, Liu et al. decomposed spatial relationships into a ternary graph representation:

  1. A fixed physical adjacency matrix (e.g., human anatomy).
  2. A learned global matrix capturing domain-invariant features.
  3. A private local matrix capturing dataset-specific variations.

This separation allows the federated model to optimize universal structural patterns across all participants while preserving local nuances without transferring site-specific topologies.

When direct node matching is impossible due to strict privacy constraints, alternative graph abstractions can act as a proxy. Zhang et al. demonstrated that item-to-item transition graphs can share global interest patterns across different applications without moving or exposing sensitive individual user profiles. Similarly, Wang et al. showed that integrating federated learning with heterogeneous graph embeddings allows multiple institutions to build high-quality global representations of complex networks that match or exceed the performance of traditional centralized models.

Vertical Graph Partitioning and Secure Alignment

Graphs can also be split vertically, meaning different organizations hold entirely different types of features or edges for the same set of nodes. For example, an e-commerce platform holds a user's purchase history graph, while a social media platform holds that same user's social circle graph.

Vertical Partitioning:
[ User Nodes ] ─── Silo 1: Purchase History Graph
               └─── Silo 2: Social Circle Graph

Mai et al. proved that vertical federated GNNs can achieve predictive accuracy within 2% of centralized models, demonstrating that the structural benefits of graphs are preserved even when the topology is split across corporate boundaries. To execute this securely, Wang et al. combined homomorphic encryption with differential privacy, allowing two parties to compute multi-party graph convolutions collaboratively without exposing their respective adjacency matrices or features.

For unsupervised tasks like network alignment (graph matching), Zhou et al. separated global model evaluation on encrypted embeddings from local client optimization. This architectural split maintains data sovereignty while allowing a global matching model to learn from aggregated structural insights across independent organizations. However, as noted by Yin et al., traditional decentralized graph embedding alignment schemes often fail strict privacy checks, necessitating explicit federated constraints where distributed workers never swap raw neighborhood structures.

Restoring Structure from Privacy Noise

To provide rigorous privacy guarantees, frameworks often inject noise via Local Differential Privacy (LDP). However, naive LDP destroys the two properties that make GNNs work: graph sparsity and feature smoothness (the tendency of connected nodes to share similar attributes).

To solve this, Lin et al. introduced mathematical calibration steps following randomized response mechanisms. These calibration steps filter out the artificial density introduced by LDP noise and smooth the perturbed features across valid edges. This ensures that a central server can still train high-quality GNNs without ever seeing the true edge topology or raw node profiles.

As these collaboratively trained models become highly valuable corporate assets, protecting them from theft becomes critical. Yang et al. extended model protection to graph domains by embedding secret trigger patterns directly into the model's decision boundaries across distributed clients. This certified watermarking process accounts for varying local topologies, ensuring ownership can be proven even if a malicious client attempts sophisticated watermark removal attacks.

Key Foundational Work

Embedding Representation of Academic Heterogeneous Information Networks Based on Federated Learning (2022) — contributed 2 concept elements to this direction.
A federated graph neural network framework for privacy-preserving personalization (2022) — contributed 2 concept elements to this direction.
Trustworthy Graph Learning: Reliability, Explainability, and Privacy Protection (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_federated_graph_learning_5316540b,
  title = {federated graph learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/federated-graph-learning/rdr_5316540b0eb44552882eb352a8edbebf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_federated_graph_learning_5316540b_v1,
  title = {federated graph learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/federated-graph-learning/rdr_5316540b0eb44552882eb352a8edbebf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake