Related Experiment Videos
FedAK: a semi-supervised one-shot framework for heterogeneous federated learning via feature-level attention-based
Hassan Salman1, Jean-François Pradat-Peyre1, Sonia Guehis2
1LIP6 UMR 7606 Sorbonne Université - CNRS & Paris Nanterre University, Paris, France.
Scientific Reports
|May 23, 2026
Summary
FedAK, a novel federated learning (FL) framework, enhances privacy by transmitting only feature representations. This method improves global model performance, especially with non-IID data.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Distributed Systems
Background:
- Federated learning (FL) allows collaborative model training while preserving data privacy.
- Key FL challenges include model heterogeneity, communication overhead, and performance issues with non-IID data.
Purpose of the Study:
- Introduce FedAK, a semi-supervised one-shot FL framework.
- Address privacy concerns and improve efficiency in FL settings.
- Reduce communication costs and data exposure.
Main Methods:
- Clients train local models and transmit only feature representations of a public dataset.
- Server uses semi-supervised aggregation with feature-level attention and knowledge distillation.
- Attention module generates ensemble features and soft pseudo-labels for unlabeled data.
Main Results:
- FedAK significantly reduces communication costs compared to traditional FL.
- The framework demonstrates superior performance against state-of-the-art one-shot FL methods.
- Effectiveness confirmed across diverse benchmark datasets, including heterogeneous and non-IID settings.
Conclusions:
- FedAK offers an efficient and privacy-preserving solution for federated learning.
- The integration of attention and knowledge distillation effectively handles non-IID data and heterogeneity.
- This framework advances the practical application of FL in real-world scenarios.