Research map / Federated learning and privacy / Personalized and heterogeneous FL
Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients
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 Topology-aware Federated Learning (TFL), which infers a client graph from model-weight similarity and uses betweenness-centrality priors to constrain distributionally robust FL for better out-of-federation generalization to unseen clients.
Problem
Federated learning models degrade sharply on unseen (out-of-federation) clients, and existing methods that generalize to unseen clients either have high communication/computation costs or poor OOF generalization.
Approach
Learn a client topology from pairwise model-weight similarity (privacy-preserving), compute betweenness-centrality-based influential-client priors, and use them to constrain a distributionally robust minimax FL objective so training balances worst-case and influential clients.
Key results
- On eICU, TFL reaches 58.41 ROC-AUC vs 57.18 for FedAvg and 57.20 for DRFA, using 2 communication rounds and 10.44M communicated parameters.
- On PACS (leave-one-domain-out, 20 clients), TFL averages 69.95% accuracy vs 68.41% for DRFA and 67.40% for FedAvg.
- With 10-cluster client clustering on eICU, wall-clock time drops from 437.61s to 133.08s (69% reduction) with ROC-AUC 57.96.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TFL instead of DRFA | Federated learning · PACS average (OOF) | accuracy: 68.41 -> 69.95 | Helped |
| use TFL instead of FedSR | Federated learning · eICU (OOF) | ROC-AUC: 57.25\textpm0.03 -> 58.41\textpm0.06 | Helped |
| use TFL instead of FedSR | Federated learning · FeTS (OOF) | DSC: 72.85\textpm {0.05} -> 74.29\textpm {0.15} | Helped |
| use TFL instead of DRFA | Federated learning · TPT-48, both OOF tasks | MSE: 0.5451 -> 0.4978; 1.8972 -> 1.7432 | Helped |
| use TFL instead of DRFA | Federated learning · TPT-48, both IF tasks | MSE: 0.1554 -> 0.1523; 0.1437 -> 0.1405 | Helped |
| use TFL instead of the strongest reported baseline | Federated learning · CIFAR-10 and CIFAR-100, 25 and 50 clients (IF) | accuracy: 70.51 -> 72.24; 64.48 -> 65.59; 38.04 -> 38.85; 37.54 -> 38.08 | Helped |
| use TFL instead of DRFA | Federated learning · OfficeHome average (OOF) | accuracy: 33.21 -> 34.89 | Helped |
| use partial model parameters instead of full parameters for similarity (ablation) | TFL · PACS (OOF) | accuracy: 69.95 -> 69.77 | No clear effect |
| vary graph sparsity threshold ε (ablation) | TFL · PACS (OOF) | accuracy: 58.61; 59.05; 58.83; 58.11; 57.73 | Mixed |
| vary topology update frequency (ablation) | TFL · PACS (OOF) | accuracy: 59.19; 58.91; 58.27; 58.80; 58.28 | No clear effect |
| vary q (ablation) | TFL · eICU (OOF) | ROC-AUC: 57.91; 58.31; 57.43; 57.36; 57.29 | Mixed |
| replace betweenness centrality with other centrality measures (ablation) | TFL · eICU (OOF) | ROC-AUC: 58.28; 57.69; 57.86; 57.57; 57.83 | Hurt |
Surprising findings
- Learned client topology from model-weight similarity identifies influential clients geographically close to the OOF region (MIDWEST and SOUTH for WEST), though the paper's own explanation is heuristic.
- Using only the last few layers for similarity gives OOF accuracy nearly identical to full-parameter similarity (69.77 vs 69.95 on PACS).
Limitations
- Client topology learning is O(N^2) in number of clients, which may be prohibitive in cross-device settings without clustering.
- Potential privacy leakage from sharing model weights used for topology inference is not fully addressed.
- High-dimensional model weights for large models increase similarity computation cost; partial-parameter workaround only tested on PACS.
Experimental setup
Datasets: eICU FeTS TPT-48 CIFAR-10 CIFAR-100 PACS OfficeHome
Benchmarks: eICU FeTS TPT-48 CIFAR-10 CIFAR-100 PACS OfficeHome
Metrics: ROC-AUC DSC MSE accuracy wall-clock time communication rounds
Baselines: FedAvg FedProx DRFA FedSR Mixup CORAL DGER DIRT-GAN
Builds on: FedAvg FedProx DRFA FedSR distributionally robust optimization
Closest papers by meaning
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.88)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.88)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.87)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.87)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.87)
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · similarity 0.87)
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.