Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedFSL-CFRD: Personalized Federated Few-Shot Learning with Collaborative Feature Representation Disentanglement
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 FedFSL-CFRD, a federated few-shot learning method that disentangles client-wise global vs. personalized features and class-wise class-relevant vs. class-irrelevant features via reconstruction, evaluated on Mini-ImageNet, Tiered-ImageNet and CIFAR-FS.
Problem
Federated few-shot learning (FedFSL) struggles with domain bias from data heterogeneity and feature coupling under sample scarcity, so personalized local models generalize poorly to unseen categories.
Approach
Decouples each client's representation into a frozen global bias feature and a local personalized feature (client-wise), and adds a few-shot-branch reconstruction branch that separates class-specific from class-irrelevant information (class-wise), within a ProtoNet/ResNet-12 federated pipeline.
Key results
- FedFSL-CFRD reaches 56.66% 5-way 1-shot and 72.15% 5-way 5-shot on Mini-ImageNet (FSL Homogeneous), above the best baseline GPFL at 56.10% and 71.41%.
- On Tiered-ImageNet (FSL Homogeneous) it reaches 60.52% and 78.95%, versus 60.14% and 78.56% for GPFL.
- The method outperforms FedFSL-Adv by roughly 1% across datasets and settings, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedFSL-CFRD instead of FedFSL-Adv | FedFSL · Mini-ImageNet, FSL Homogeneous, 5-way 1-shot | accuracy: 55.75±1.65 -> 56.66±1.52 | Helped |
| Use FedFSL-CFRD instead of FedFSL-Adv | FedFSL · Mini-ImageNet, FSL Heterogeneous, 5-way 1-shot | accuracy: 52.52±1.68 -> 53.89±1.06 | Helped |
| Use FedFSL-CFRD instead of FedFSL-Adv | FedFSL · Tiered-ImageNet, FSL Homogeneous, 5-way 5-shot | accuracy: 78.19±1.38 -> 78.95±1.28 | Helped |
| Use FedFSL-CFRD instead of FedFSL-Adv | FedFSL · CIFAR-FS, FSL Heterogeneous, 5-way 5-shot | accuracy: 72.00±1.80 -> 72.44±0.96 | Helped |
| Add client-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Mini-ImageNet, FSL Homogeneous, 5-way 1-shot | accuracy: 51.32±1.41 -> 51.96±1.75 | Helped |
| Add client-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Tiered-ImageNet, FSL Heterogeneous, 5-way 5-shot | accuracy: 71.25±1.61 -> 72.84±1.75 | Helped |
| Add class-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Mini-ImageNet, FSL Homogeneous, 5-way 1-shot | accuracy: 51.32±1.41 -> 53.29±1.68 | Helped |
| Add class-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Tiered-ImageNet, FSL Heterogeneous, 5-way 5-shot | accuracy: 71.25±1.61 -> 75.91±1.54 | Helped |
| Add both client-wise and class-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Mini-ImageNet, FSL Homogeneous, 5-way 5-shot | accuracy: 66.67±2.01 -> 72.15±0.98 | Helped |
| Add both client-wise and class-wise feature disentanglement (ablation) | ProtoNet FedFSL framework · Tiered-ImageNet, FSL Heterogeneous, 5-way 1-shot | accuracy: 53.88±1.41 -> 58.04±1.00 | Helped |
Limitations
- Gains over the strongest baselines are small (often about 0.5-1 point) with overlapping confidence intervals.
- Performance decreases as the number of clients grows, especially under heterogeneous data.
- Evaluated only on ResNet-12 with ProtoNet and on three image benchmarks.
Experimental setup
Datasets: Mini-ImageNet Tiered-ImageNet CIFAR-FS
Benchmarks: Mini-ImageNet Tiered-ImageNet CIFAR-FS
Metrics: 5-way 1-shot accuracy 5-way 5-shot accuracy
Baselines: FedAvg FedProx FedProto PerFedAvg pFedMe FedFomo FedALA GPFL FedFSL-Adv
Builds on: ProtoNet FedAvg FedALA GPFL MOON
Closest papers by meaning
- Personalized Federated Learning via Feature Distribution Adaptation (NeurIPS 2024 · similarity 0.87)
- Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation (NeurIPS 2025 · similarity 0.86)
- Flexible Sharpness-Aware Personalized Federated Learning (AAAI 2025 · similarity 0.85)
- FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data (ICLR 2024 · similarity 0.85)
- Causally Motivated Personalized Federated Invariant Learning with Shortcut-Averse Information-Theoretic Regularization (ICML 2024 · similarity 0.85)
- pFedES: Generalized Proxy Feature Extractor Sharing for Model Heterogeneous Personalized Federated Learning (AAAI 2025 · similarity 0.85)
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.