Research map / Federated learning and privacy / Personalized and heterogeneous FL

GHOST: Generalizable One-Shot Federated Graph Learning with Proxy-Based Topology Knowledge Retention

ICML 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated semi-supervised and graph learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GHOST, a one-shot federated graph learning framework that trains a per-client proxy generator with dual-level (feature and topology) alignment, then trains a global GNN on synthetic pseudo graphs while protecting topology-critical parameters against forgetting.

Problem

Federated graph learning under a single communication round suffers from poor generalization of the global model across diverse graph structures and catastrophic forgetting of topological knowledge under extreme structural heterogeneity.

Approach

Each client trains a proxy model that generates pseudo feature-graphs from Gaussian noise and labels, aligned to local data via a divergence, dispersion, and fused Gromov-Wasserstein loss; the server samples pseudo graphs from all proxies to train a global GNN, with a Topology-Consistency Criterion that identifies parameters important for structure and regularizes them via a knowledge retention lo

Key results

• GHOST achieves 50.41% on Cora vs. 30.61% for FedAvg and 30.89% for the best baseline FedCVAE in the high-heterogeneity setting (alpha=0.05), and is best on all seven datasets.
• Ablation on Cora: DLA alone raises accuracy from 12.65% to 46.93%, and adding TCKR yields 50.41%.
• Privacy experiments show shifted-center and uniform-guess attacks cause a 50%-80% accuracy drop for an eavesdropper.

Measured effects reported in the paper

Surprising findings

• Existing FGL methods such as FedPub and FedTAD fall below FedAvg on many datasets in the one-shot setting, and FedTAD runs out of memory on Ogbn-Arxiv.
• Several traditional FL baselines (e.g., FedNova, FedRCL) collapse to near-random accuracy under this one-shot setting.

Limitations

• Noise in local data may interfere with proxy model learning of local-specific knowledge.
• Hyperparameter scales (e.g., lambda_f set to 1e-4 and 1e-7 on some datasets) require manual tuning.
• Evaluated only on node classification with a two-layer GCN backbone and 10 clients in the main experiments.

Experimental setup

Datasets: Cora CiteSeer PubMed Chameleon Amazon-Photo Coauthor-CS Ogbn-Arxiv

Benchmarks: Cora CiteSeer PubMed Chameleon Amazon-Photo Coauthor-CS Ogbn-Arxiv

Metrics: node classification accuracy

Baselines: FedAvg FedProx FedNova FedRCL FedPub FedTAD FedGTA DENSE FedCVAE FedSD2C

Builds on: FedAvg Elastic weight consolidation-style regularization (Zenke et al., 2017) DENSE FedCVAE

Closest papers by meaning

• Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.90)
• $S^2$FGL: Spatial Spectral Federated Graph Learning (ICML 2025 · similarity 0.89)
• FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.88)
• FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding Decorrelation (AAAI 2025 · similarity 0.88)
• Virtual Nodes Can Help: Tackling Distribution Shifts in Federated Graph Learning (AAAI 2025 · similarity 0.88)
• Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.88)