A Federated Parameter Aggregation Method for Node Classification Tasks with Different Graph Network Structures

TL;DR

Addresses node classification under horizontal federated learning with heterogeneous graph structures. Proposes FLGNN, a parameter-aggregation method that shares multi-layer GNN weights $W$ across clients and aggregates per-layer parameters via FedAvg, and FLGNN+ which assigns dynamic weights $\gamma_{u,t}^n$ to account for different edge types; privacy defenses via membership inference attacks and differential privacy noise drawn from $Lap(\Delta f/\epsilon)$ are evaluated. Demonstrates that FLGNN achieves near-centralized performance (within $1\%$–$2\%$) on real datasets and FLGNN+ improves robustness to edge-type heterogeneity, while providing measurable privacy protection. Collectively, the work enables privacy-preserving collaboration across diverse graph networks, offering tunable privacy-utility trade-offs for practical deployment.

Abstract

Over the past few years, federated learning has become widely used in various classical machine learning fields because of its collaborative ability to train data from multiple sources without compromising privacy. However, in the area of graph neural networks, the nodes and network structures of graphs held by clients are different in many practical applications, and the aggregation method that directly shares model gradients cannot be directly applied to this scenario. Therefore, this work proposes a federated aggregation method FLGNN applied to various graph federation scenarios and investigates the aggregation effect of parameter sharing at each layer of the graph neural network model. The effectiveness of the federated aggregation method FLGNN is verified by experiments on real datasets. Additionally, for the privacy security of FLGNN, this paper designs membership inference attack experiments and differential privacy defense experiments. The results show that FLGNN performs good robustness, and the success rate of privacy theft is further reduced by adding differential privacy defense methods.

Figures (6)

• Figure 1: Three federation situations are depicted in the diagram. Two clients' networks are represented by A and B. The blue nodes show common nodes between the two networks, while the red and yellow nodes represent network-specific nodes. Different sorts of edges are shown by the dotted and solid lines.
• Figure 2: The left side of the figure depicts the process of federated training and aggregation, where $[\![*]\!]$ represents the encrypted parameter, $i$ is the number of graph neural network layers, and $n,n_2,n_3$ is the number of feature dimension. The right side of the figure depicts FLGNN specific structure. Furthermore, $m$ represents the number of categories, and $\hbox{\boldmath$a$}$ represents the head of the graph attention mechanism.
• Figure 3: Federal training flow chart.
• Figure 4: Trend chart of the influence of the number of clients on federated aggregation. The abscissa is the number of clients, and the ordinate is the average accuracy of Full_Client minus individual training and minus FLGNN
• Figure 5: Trend chart of the influence of the number of clients on federated aggregation. The abscissa is the number of clients, and the ordinate is the average accuracy of Full_Client minus individual training and minus FLGNN