Research map / Federated learning and privacy / Personalized and heterogeneous FL
FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts
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 FLUX, a hard clustered federated learning framework that clusters clients via privacy-preserving latent-moment descriptors and an adaptive DBSCAN, handling four distribution shifts and assigning unseen unlabeled clients to cluster models at test time.
Problem
Federated learning under non-IID client data suffers accuracy loss; existing clustered and personalized FL methods handle only some shift types, need the number of clusters in advance, have high overhead, and cannot assign unseen unlabeled clients at test time.
Approach
Clients compute a compact descriptor from the first two moments of shared-PCA-reduced encoder latents (marginal and class-conditional), which approximates the 2-Wasserstein distance between client distributions; the server clusters descriptors with an adaptive DBSCAN, and test clients are assigned to the nearest centroid using only label-agnostic marginal moments.
Key results
- Average accuracy gains of up to 23.0 pp on FMNIST and 12.4 pp on MNIST over best baselines under known association setting, per the paper's robustness analysis.
- In the test-phase setting, FLUX improves accuracy by up to 7.84 pp (MNIST), 11.86 pp (FMNIST), 2.17 pp (CIFAR-10), and 1.5 pp (CIFAR-100) over the best baselines.
- On MNIST with 100 clients, FLUX sustains accuracy above 84% in the test phase while APFL drops to over 70%; the fastest CFL baseline FeSEM takes more than 4x FLUX's training time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Flux instead of the strongest baseline for test-time client assignment | Federated learning · MNIST, test phase; averaged across heterogeneity levels | accuracy: 7.84 pp | Helped |
| Use Flux instead of the strongest baseline for test-time client assignment | Federated learning · FMNIST, test phase; averaged across heterogeneity levels | accuracy: 11.86 pp | Helped |
| Use Flux instead of the strongest baseline for test-time client assignment | Federated learning · CIFAR-10, test phase; averaged across heterogeneity levels | accuracy: 2.17 pp | Helped |
| Use Flux instead of the strongest baseline for test-time client assignment | Federated learning · CIFAR-100, test phase; averaged across heterogeneity levels | accuracy: 1.5 pp | Helped |
| Use Flux instead of the strongest baseline | Federated learning · CheXpert, known association | ROC AUC: up to 8.5 pp | Helped |
| Use Flux instead of the strongest baseline | Federated learning · CheXpert, test phase | ROC AUC: up to 17.5 pp | Helped |
| Use Flux instead of the strongest baseline | Federated learning · Office-Home, test phase | accuracy: up to 1.3 pp | Helped |
| Use Flux instead of the strongest baseline | Federated learning · MNIST, FMNIST, CIFAR-10, and CIFAR-100; known association | accuracy: 12.4, 23.0, 7.0, and 3.0 percentage points | Mixed |
Surprising findings
- Most metric- and parameter-based CFL methods collapse to a single global model on Office-Home, failing to capture subtle real-world shifts.
- ATP, an unsupervised test-time adaptation method, shows instability and overconfident incorrect predictions on CIFAR-100.
Limitations
- Requires a sufficient amount of diverse training data to obtain statistically robust descriptors.
- One-shot clustering; does not handle distribution drift over time.
- P(Y|X) concept shift cannot be resolved at test time because test-time descriptors use only marginal features.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 CIFAR-100 CheXpert Office-Home
Benchmarks: MNIST Fashion-MNIST CIFAR-10 CIFAR-100 CheXpert Office-Home
Metrics: accuracy ROC AUC training time communication overhead
Baselines: FedAvg CFL IFCA FeSEM FedEM FedRC FedDrift pFedMe APFL ATP
Builds on: FedAvg IFCA DBSCAN ANDA
Closest papers by meaning
- Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies (ICLR 2025 · similarity 0.91)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.90)
- Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.89)
- FedImpro: Measuring and Improving Client Update in Federated Learning (ICLR 2024 · similarity 0.89)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.89)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.89)
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.