Research map / Federated learning and privacy / Privacy attacks and secure learning
An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective
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 ESFMC, a federated multi-view clustering method that splits client features via an information-theoretic objective so only clustering-related features are shared, keeping sample-related features local, with a collaborative alignment extension for incomplete data.
Problem
Federated multi-view clustering methods share model parameters or intermediate features to handle cross-client view heterogeneity, but this sharing exposes clients to privacy attacks such as reconstruction and model inversion, creating a trade-off between privacy and clustering performance.
Approach
Split each client's latent features into clustering-related features (shared with the server) and sample-related features (kept local) using mutual-information objectives, train local models via SGLD-based Gibbs posterior with a model-data mutual information constraint, and align non-overlapping samples through shared clustering assignments in incomplete settings.
Key results
- ESFMC reaches 91.50 ACC on Caltech versus 90.14 for the next best method (AGDIMC) in general scenarios.
- On Synthetic3d, ESFMC achieves 97.32 ACC and 91.23 ARI, well above the second-best FedMVFCM (93.07 ACC) and DSIMVC (85.31 ARI).
- In incomplete scenarios on BDGP (overlap 0.5), ESFMC gets 94.9 ACC and 87.8 ARI, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ESFMC instead of AGDIMC | Federated multi-view clustering · Caltech | ACC: 90.14 -> 91.50 | Helped |
| use ESFMC instead of AGDIMC | Federated multi-view clustering · Caltech | NMI: 82.91 -> 84.54 | Helped |
| use ESFMC instead of AGDIMC | Federated multi-view clustering · Caltech | ARI: 80.27 -> 83.45 | Helped |
| use ESFMC instead of FCUIF | Federated multi-view clustering · HW | ACC: 97.85 -> 98.02 | Helped |
| use ESFMC instead of FCUIF | Federated multi-view clustering · HW | NMI: 95.16 -> 95.34 | Helped |
| use ESFMC instead of FCUIF | Federated multi-view clustering · HW | ARI: 95.30 -> 95.57 | Helped |
| use ESFMC instead of DSIMVC | Federated multi-view clustering · MNIST-USPS | ACC: 98.08 -> 97.24 | Hurt |
| use ESFMC instead of DSIMVC | Federated multi-view clustering · MNIST-USPS | NMI: 94.81 -> 93.19 | Hurt |
| use ESFMC instead of DSIMVC | Federated multi-view clustering · MNIST-USPS | ARI: 95.05 -> 94.02 | Hurt |
| use ESFMC instead of DSIMVC | Federated multi-view clustering · Synthetic3d | ACC: 94.83 -> 97.32 | Helped |
| use ESFMC instead of DSIMVC | Federated multi-view clustering · Synthetic3d | NMI: 81.10 -> 87.67 | Helped |
| use ESFMC instead of FedMVFCM | Federated multi-view clustering · Synthetic3d | ARI: 87.03 -> 91.23 | Helped |
Surprising findings
- Removing the sample-related reconstruction loss (w/o L1) or the mutual-information minimization loss (w/o L3) has minimal effect on clustering performance in the ablation.
- On Synthetic3d, ESFMC's ACC is slightly lower than FedMVFCM's on some metrics; the paper notes FedMVFCM's second-best status there.
Limitations
- Designed only for semi-honest participants; malicious attacks are not covered.
- Privacy is validated only with a single model inversion attack on MNIST-USPS.
- Incomplete scenarios may require sharing clustering assignments, increasing exposure.
Experimental setup
Datasets: Caltech HW MNIST-USPS Synthetic3d BDGP Scene
Benchmarks: Caltech HW MNIST-USPS Synthetic3d BDGP Scene
Metrics: ACC NMI ARI
Baselines: GIMC-FLSD HCP-IMSC IMVC-CBG DSIMVC AGDIMC FedDMVC FedMVFCM FedMVFPC FCUIF
Builds on: information bottleneck FedDMVC Gibbs posterior / SGLD K-means
Closest papers by meaning
- Efficient Federated Incomplete Multi-View Clustering (ICML 2025 · similarity 0.93)
- Scalable Federated One-Step Multi-View Clustering with Tensorized Regularization (AAAI 2025 · similarity 0.89)
- Graph Consistency and Diversity Measurement for Federated Multi-View Clustering (AAAI 2025 · similarity 0.89)
- OmniFC: Rethinking Federated Clustering via Lossless and Secure Distance Reconstruction (NeurIPS 2025 · similarity 0.88)
- Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views (NeurIPS 2024 · similarity 0.87)
- Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance (ICML 2025 · similarity 0.86)
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.