Federated Graph Neural Network

Each client holds a local subgraph containing a subset of nodes and edges from a global graph. The raw topology and node features never leave the client device. Training occurs locally on this partial view, and only model gradients or embedding updates are transmitted to the aggregation server. This architecture is critical for scenarios like cross-institutional drug discovery, where molecular interaction graphs are proprietary.

A core challenge in FedGNNs is that a node's k-hop neighborhood may span across multiple clients. To address this, techniques like federated neighbor sampling and embedding propagation are used. Clients may request anonymized embeddings of missing neighbors from the server or other clients, enabling accurate local message passing without violating data locality. This preserves the message-passing integrity of the GNN.

FedGNNs operate in two primary modes:

• Graph-Level Federation: Each client owns a set of independent graphs (e.g., molecular graphs). Models are trained to predict graph-level properties, and only readout layer gradients are aggregated.
• Node-Level Federation: Clients hold overlapping subgraphs of one large graph (e.g., a social network). Training requires careful handling of cross-edge dependencies to avoid information leakage.

Unlike standard federated learning, FedGNNs must contend with topological non-IIDness. Local subgraphs can have vastly different degree distributions, community structures, and label densities. Specialized aggregation strategies like FedSage+ and structure-aware personalization layers are employed to prevent the global model from overfitting to the dominant topological patterns of a few clients.

FedGNNs enable collaborative link prediction across data silos. For example, two banks can jointly train a model to detect fraudulent transaction patterns without sharing their customer graphs. Techniques like secure aggregation and differential privacy are integrated into the gradient sharing process to prevent inference attacks that could reconstruct sensitive edge relationships from model updates.

To ensure a coherent global representation, FedGNNs often employ contrastive learning objectives. A local node embedding is pulled closer to its corresponding global embedding while being pushed away from other nodes. This alignment step, performed during aggregation, mitigates client drift caused by the disparate local graph structures and accelerates convergence of the global GNN model.

A highly expressive GNN architecture that generalizes the Weisfeiler-Lehman test for graph isomorphism. In a federated setting, GIN's sum aggregation over local node neighborhoods makes it theoretically suitable for distinguishing subgraph structures across different clients, ensuring that the global model can capture distinct local topological patterns without sharing raw adjacency data.

The core challenge in federated graph learning where local subgraphs on different clients exhibit statistical heterogeneity in node degree distribution, feature space, and label density. Unlike standard non-IID data, graph non-IID involves structural skew, where one client may hold a dense community cluster while another holds a sparse periphery, causing severe local model drift during aggregation.

A cryptographic protocol that allows a central server to compute the sum of encrypted adjacency-sensitive model updates without inspecting individual graph topologies. This is critical for cross-silo federated learning scenarios, such as inter-bank fraud detection, where the transaction graph structure itself is proprietary and must remain hidden even during the aggregation process.

An extension of the inductive GraphSAGE framework to a decentralized setting, enabling the generation of embeddings for unseen nodes. In this paradigm, clients locally sample and aggregate features from a node's neighborhood, sharing only the learned aggregation function weights. This allows the global model to generalize to new subgraph structures without requiring retraining on a centralized graph.

A privacy-preserving technique to combat the isolation of local subgraphs. Instead of sharing raw edges, clients exchange synthetic node embeddings or anonymized motif counts generated by a differential privacy mechanism. This provides the local GNN with a statistically similar global context, mitigating the catastrophic forgetting of long-range dependencies that occurs when training only on isolated local ego-networks.

A decentralized implementation of attention mechanisms on graphs where clients compute local attention coefficients. The challenge lies in cross-client attention normalization, as the softmax function is typically global. Federated GAT variants often use local normalization or share attention coefficient statistics to ensure that the importance a node assigns to its neighbors is calibrated across the entire distributed system.