Rethinking Federated Graph Learning: A Data Condensation Perspective

TL;DR

This paper tackles subgraph heterogeneity in Federated Graph Learning by introducing FedGM, a condensation-based paradigm that replaces conventional model-parameter transfers with a condensed graph as the optimization carrier. FedGM operates in two stages: (1) local subgraph condensation via one-shot gradient matching to produce compact, informative subgraphs, and (2) federated gradient matching to refine a global condensed graph using class-wise gradients, enabling robust global GNN training on the condensed structure. Empirically, FedGM outperforms multiple baselines across six datasets, with stronger gains as client count increases and reduced communication and privacy risk due to single-shot data transmission. The approach demonstrates that condensed graphs can effectively summarize distributed graph knowledge, offering a practical, scalable route for privacy-preserving FGL with improved convergence properties.

Abstract

Federated graph learning is a widely recognized technique that promotes collaborative training of graph neural networks (GNNs) by multi-client graphs.However, existing approaches heavily rely on the communication of model parameters or gradients for federated optimization and fail to adequately address the data heterogeneity introduced by intricate and diverse graph distributions. Although some methods attempt to share additional messages among the server and clients to improve federated convergence during communication, they introduce significant privacy risks and increase communication overhead. To address these issues, we introduce the concept of a condensed graph as a novel optimization carrier to address FGL data heterogeneity and propose a new FGL paradigm called FedGM. Specifically, we utilize a generalized condensation graph consensus to aggregate comprehensive knowledge from distributed graphs, while minimizing communication costs and privacy risks through a single transmission of the condensed data. Extensive experiments on six public datasets consistently demonstrate the superiority of FedGM over state-of-the-art baselines, highlighting its potential for a novel FGL paradigm.

Figures (6)

• Figure 1: (a) The conventional subgraph-FL framework in the subgraph heterogeneity scenario where node colors represent different labels. (b) The condensation-based subgraph-FL framework, which trains a robust global model by integrating condensed knowledge.
• Figure 2: (a) Label distribution based on random data split, where the color gradient from white to blue indicates the increasing number of nodes held by different clients in each class. (b) Optimization performance of various methods under subgraph heterogeneity scenarios. The x-axis of the line plot represents federated training rounds, with "Local" indicating model performance in siloed settings.
• Figure 3: Overview of our proposed FedGM paradigm. We first perform local subgraph condensation and the central server integrates the condensed subgraphs. Subsequently, the server receives class-wise gradients from real subgraphs in federated communication to enhance the quality of condensed knowledge. Ultimately, the global model is trained on the condensed graph and then distributed to the clients.
• Figure 4: This is the representation of the local and global gradient matching in the model parameter space. The gradient matching iteratively optimizes the condensed data by minimizing the distance between gradients generated by the real and condensed data on the model, ultimately aligning the low-loss region of the condensed data within the low-loss region of the real data. The blue intersecting region in the right panel represents shared intra-class knowledge.
• Figure 5: Performance of FedGM with different numbers of clients.