Research map / Federated learning and privacy / Privacy attacks and secure learning
Federated Graph Condensation with Information Bottleneck Principles
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 FedGC, a federated graph condensation framework that matches class-weighted aggregated client gradients and uses a local information-bottleneck feature transformation to resist membership inference on the released condensed graph.
Problem
Graph condensation assumes centralized data, and condensed graphs released in a federated setting can leak membership of local training nodes via membership inference attacks.
Approach
Decouple gradient matching into client-side gradient computation and server-side matching with class-aware weighted aggregation, plus a local IB-based node feature transformation (structure kept fixed) that reduces mutual information between original and transformed features, with one-step encrypted neighbor aggregation and federated self-training to label nodes and align feature spaces.
Key results
- On Ogbn-arxiv at r=0.25%, FedGC reaches 75.8% accuracy versus 70.8% for FedGCN and 66.1% for SFGC.
- On Cora at 2.6%, the MIA AUC drops to 53.2% for FedGC versus 76.4% without defense, while accuracy stays at 80.8%.
- Removing IB yields slightly lower accuracy on most datasets and a higher MIA AUC on Citeseer (55.0 with FedGC vs 87.2 without defense).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local graph transformation with information bottleneck (ablation) | FedGC · Cora, 2.6% | MIA AUC (%): 76.4 -> 53.2 | Helped |
| add local graph transformation with information bottleneck (ablation) | FedGC · Citeseer, 1.8% | MIA AUC (%): 87.2 -> 55.0 | Helped |
| add local graph transformation with information bottleneck (ablation) | FedGC · Ogbn-arxiv, 0.25% | MIA AUC (%): 64.0 -> 51.6 | Helped |
| add local graph transformation with information bottleneck (ablation) | FedGC · Flickr, 0.5% | MIA AUC (%): 62.1 -> 54.8 | Helped |
| add local graph transformation with information bottleneck (ablation) | FedGC · Reddit, 0.1% | MIA AUC (%): 61.1 -> 52.3 | Helped |
| add one-step communication to recover cross-client neighbor aggregations (ablation) | FedGC · Cora, Citeseer, and Ogbn-arxiv condensed-graph accuracy | accuracy | Helped |
| add self-training to label local unlabeled nodes (ablation) | FedGC · Cora and Citeseer condensed-graph accuracy | accuracy | Helped |
| add local graph transformation with information bottleneck (ablation) | FedGC · condensed-graph accuracy across almost all datasets | accuracy | Helped |
| add local fine-tuning with client data (ablation) | FedGC · personalized local evaluation | accuracy | Helped |
| use FedGC instead of centralized SFGC | condensed-graph node classification · Cora, 2.6% | accuracy (%): 81.7${\pm}$0.5 -> 80.8${\pm}$1.8 | Hurt |
| use FedGC instead of centralized SFGC | condensed-graph node classification · Citeseer, 1.8% | accuracy (%): 72.4{\pm}0.4 -> ${70.5}{\pm}1.9 | Hurt |
| use FedGC instead of FedGCN | condensed-graph node classification · Ogbn-arxiv, 0.25% | accuracy (%): $70.8{\pm}1.7$ -> 75.8${\pm}$0.2 | Helped |
Surprising findings
- FedGC outperforms FGL methods in the non-i.i.d. setting, which the authors attribute to the condensed graph capturing common knowledge across subgraphs.
- Model-based defenses (Reg, LDP) fail to reduce MIA on the public condensed graph, while the feature-level IB transformation substantially lowers AUC.
Limitations
- Self-training in the condensation phase adds communication and computation cost, which the authors leave as future work.
- MIA performance rises with the number of clients on Cora, suggesting that more clients can increase leakage.
- Only node classification on five datasets with GCN-based backbones is evaluated.
Experimental setup
Datasets: Cora Citeseer Ogbn-arxiv Flickr Reddit
Benchmarks: Cora Citeseer Ogbn-arxiv Flickr Reddit
Metrics: accuracy AUC of membership inference attack
Baselines: GCOND DosCond SFGC FedGCN FedSage+ BDS-GCN PL Reg LDP GCN
Builds on: GCOND FedGCN information bottleneck DosCond FedAvg
Closest papers by meaning
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.89)
- Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.89)
- Federated Graph-Level Clustering Network (AAAI 2025 · similarity 0.88)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.88)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.88)
- GRAIN: Exact Graph Reconstruction from Gradients (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.