Abstract
Federated few-shot learning (FedFSL) aims to enable clients to obtain personalized generalization models for unseen categories with only a small number of referenceable samples in a distributed collaborative training paradigm. Most existing FedFSL-related algorithms suffer from domain bias and feature coupling in the presence of data heterogeneity and sample scarcity. In this work, we propose a collaborative feature representation disentanglement (CFRD) scheme for FedFSL to address these issues. After each client receives the global aggregation parameters, the original feature representation is decoupled into global communal features and local personality features with personalized bias representation, maintaining both global consistency and local relevance. On the few-shot metric space, category-independent information is encoded by class-specific and class-irrelevant reconstructions to separate discriminative features. The proposed scheme collaboratively accomplishes global domain bias feature disentanglement and local category degradation feature disentanglement from client-wise and class-wise perspectives. Experiments on three few-shot benchmark datasets demonstrate that our proposed method outperforms state-of-the-art approaches in both global generality and local specificity.
Introduction
Federated Learning (FL) allows multiple parties to train a model collaboratively without sharing private data. However, when combined with Few-Shot Learning (FSL)—where models must recognize new categories from just one or five examples—two major hurdles emerge: Domain Bias and Feature Coupling.
Domain bias occurs because data is distributed across different clients with varying environments (e.g., lighting or camera types), making it difficult for a global model to find a universal representation. Feature coupling refers to the difficulty of distinguishing between category-defining features (e.g., the "dog-ness" of a dog) and irrelevant background noise (e.g., grass) when data is scarce. FedFSL-CFRD addresses these by explicitly separating these conflicting signals.
Methodology: The CFRD Framework
The core of the proposed framework is a dual-layer feature disentanglement strategy designed to "clean" the latent space and preserve local relevance without sacrificing global knowledge.
1. Client-Wise Disentanglement
Upon receiving the global model, a client decouples the feature representation into global communal features (shared knowledge) and local personality features (client-specific bias). The server aggregates these local biases into a "global bias feature" to coordinate training while maintaining local specificity.
2. Class-Wise Disentanglement
Within the few-shot metric space, the model further splits features into class-specific and class-irrelevant components. To enforce this, a reconstruction branch uses a decoder to rebuild the original features from both branches. This process forces the model to relegate non-discriminative information to the irrelevant branch, leaving the classification branch focused on pure, discriminative traits.
Experimental Setup
The authors evaluated FedFSL-CFRD on three benchmark datasets: Mini-ImageNet, Tiered-ImageNet, and CIFAR-FS. Experiments were conducted under both homogeneous (shared categories) and heterogeneous (split categories) settings, with the latter providing a more realistic test of federated systems.
Key Results and Discussion
FedFSL-CFRD was compared against nine state-of-the-art methods, including traditional FL (FedAvg, FedProx) and specialized personalized/few-shot FL algorithms (FedProto, FedALA, GPFL).
- Superior Accuracy: In the 5-way 1-shot heterogeneous setting on Tiered-ImageNet, CFRD achieved 58.04% accuracy, significantly outperforming competitors. In 5-shot scenarios, accuracy reached 78.95%.
- Faster Convergence: As shown in the performance curves across 10 clients, CFRD reaches stable, high accuracy in fewer communication rounds than its peers, which is critical for reducing network costs in federated systems.
- Robust Feature Clustering: t-SNE visualizations (not shown) confirm that CFRD produces tighter, more distinct clusters for different categories compared to standard FedFSL approaches.
Conclusion
FedFSL-CFRD effectively mitigates the challenges of data heterogeneity and sample scarcity through collaborative disentanglement. By treating client-specific bias as a "personality" and using reconstruction to filter noise, the framework achieves state-of-the-art performance. Future work may focus on reducing the computational overhead introduced by the additional decoder branches to better support resource-constrained edge devices.
Summary updated