Research map / Federated learning and privacy / Personalized and heterogeneous FL
Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies
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 HCFL, a four-tier unified framework for clustered federated learning that subsumes existing methods, and HCFL+, which adds sample-wise soft clustering weights, feature-extractor/predictor split, adaptive soft clustering, and fine-grained prototype-based distance metrics.
Problem
Clustered federated learning methods use diverse learning frameworks that are hard to integrate; existing methods also struggle with intra-client inconsistent clustering weights, communication/computation efficiency, adaptive cluster-number selection for soft clustering, and distance metrics that do
Approach
A unified dual-variable objective and four-tier procedure (cluster formulation, cluster weight calculation, adaptive cluster procedure, distance metric) that lets existing components be recombined; HCFL+ uses sample-wise weights for training, client-wise weights for testing, a shared feature extractor with cluster-specific predictors, split/merge adaptive clustering for soft clustering, and protot
Key results
- On CIFAR10 (beta=0.2), HCFL+ (FedRC) reaches 69.67% test accuracy with 3 clusters, versus 59.41% for FedRC with fixed K=3.
- On CIFAR100 (beta=0.2), HCFL+ (FedRC) achieves 36.50% test accuracy, close to the best baseline FedRC at 37.62%.
- HCFL+ (FedRC) on Tiny-ImageNet reaches 40.65% test accuracy, above FedRC's 39.73%, with 3 final clusters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Extend FedRC with HCFL+ adaptive clustering and feature extractor-classifier splitting | FedRC · CIFAR10, β = 0.2 | Test accuracy: 59.41 ± 0.19 -> 69.67 ± 1.27 | Helped |
| Extend FedRC with HCFL+ adaptive clustering and feature extractor-classifier splitting | FedRC · CIFAR100, β = 0.2 | Test accuracy: 37.62 ± 0.16 -> 36.50 ± 0.28 | Hurt |
| Extend FedRC with HCFL+ adaptive clustering and feature extractor-classifier splitting | FedRC · Tiny-Imagenet, β = 0.2 | Test accuracy: 39.73 ± 0.04 -> 40.65 ± 0.36 | Helped |
| Lower ρ from 0.3 to 0.05 (ablation) | HCFL+ (FedEM) · CIFAR100, β = 0.2 | Validation accuracy: 50.72 ± 2.97 -> 60.36 ± 1.47 | Mixed |
| Lower ρ from 0.3 to 0.05 (ablation) | HCFL+ (FedEM) · CIFAR100, β = 0.2 | Test accuracy: 32.13 ± 0.18 -> 22.95 ± 1.44 | Hurt |
| Use smaller ρ (ablation) | HCFL+ · CIFAR100 | Validation accuracy: 69.25 -> 78.98 | Helped |
| Use smaller ρ (ablation) | HCFL+ · CIFAR100 | Test accuracy: 36.50 -> 14.33 | Hurt |
| Remove the loss-confidence factors from the client distance metric (ablation) | HCFL+ (FedRC) · CIFAR10, β = 0.2 | Test accuracy: 69.67 ± 1.27 -> 56.14 ± 8.11 | Hurt |
| Use gradient similarity instead of feature-prototype distances (ablation) | HCFL+ (FedRC) · CIFAR10, β = 0.2 | Test accuracy: 69.67 ± 1.27 -> 67.83 ± 1.70 | Hurt |
| Use mean rather than maximum class distance (ablation) | HCFL+ (FedRC) · CIFAR10, β = 0.2 | Test accuracy: 69.67 ± 1.27 -> 68.82 ± 0.59 | Hurt |
| Remove the loss-confidence factors from the client distance metric (ablation) | HCFL+ (FedRC) · CIFAR100, β = 0.2 | Test accuracy: 36.50 ± 0.28 -> 34.11 ± 1.58 | Hurt |
| Use mean rather than maximum class distance (ablation) | HCFL+ (FedRC) · CIFAR100, β = 0.2 | Test accuracy: 36.50 ± 0.28 -> 35.97 ± 0.80 | Hurt |
Surprising findings
- Baselines such as ICFL and StoCFL achieve higher validation accuracy only by splitting clients into near-individual clusters (K near 100), which overfits and yields much lower test accuracy.
- FedRC-based HCFL+ prefers larger mu-tilde (more sample-wise weight difference), whereas FedEM-based HCFL+ prefers smaller values.
Limitations
- Only a few typical component choices per tier were tested; broader choices are left unexplored.
- Experiments are limited to image classification with MobileNet-V2 and ResNet18 on CIFAR and Tiny-ImageNet.
- Some HCFL+ variants trade test accuracy for validation accuracy depending on the splitting threshold rho.
Experimental setup
Datasets: CIFAR10 CIFAR100 Tiny-ImageNet CIFAR10-C CIFAR100-C
Benchmarks: CIFAR10 CIFAR100 Tiny-ImageNet
Metrics: validation accuracy test accuracy number of clusters K^T
Baselines: FedAvg FeSEM IFCA FedEM FedRC CFL ICFL StoCFL
Builds on: FedEM FedRC IFCA FeSEM CFL
Closest papers by meaning
- Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.92)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.91)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.91)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.88)
- Asynchronous Federated Clustering with Unknown Number of Clusters (AAAI 2025 · similarity 0.88)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · 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.