Randomized trial demonstrates improved global learning efficiency in heterogeneous federated learning, suggesting a novel approach to privacy preservation.
Key Points
This study aims to address challenges in federated learning, particularly model heterogeneity and communication overhead, through the FedAK framework.
Developed FedAK, a semi-supervised one-shot FL framework integrating feature-level attention and knowledge distillation.
Clients train local models on private labeled data and send feature representations of a public dataset to the server.
Employed a semi-supervised aggregation strategy using an attention-based module to generate pseudo-labels for a global model.
FedAK consistently outperforms four state-of-the-art one-shot FL methods across four benchmark datasets.
Demonstrated efficiency under heterogeneous and non-IID conditions.
Achieved significant reduction in communication costs by transmitting only feature representations.