An adaptive federated few-shot learning method with intelligent device selection

Abstract

Federated learning, which enables the training of intelligent models without transferring data and solely by relying on the computational capabilities of distributed devices, faces significant challenges in few-shot scenarios due to data heterogeneity, resource constraints, and slow convergence. The primary obstacle lies in the optimal selection of participating devices, as suboptimal choices can reduce model accuracy and substantially increase latency. In this study, we introduce AdaptFFSL-DS, an adaptive decision-making framework for Federated Few-Shot Learning. By integrating ResFed as the local model with an intelligent device-selection agent, this method first evaluates the system-level and statistical characteristics of each device and then selects an appropriate subset of devices for each learning round. Furthermore, by adaptively adjusting the number of local epochs, the framework maintains an effective balance between accuracy and latency. The experimental results demonstrate that the proposed approach reduces the estimated aggregate device-latency by nearly one-third without a notable loss in accuracy and delivers up to 11.88% higher accuracy compared to intelligently tuned FedProx. Moreover, AdaptFFSL-DS remains robust under various forms of heterogeneity, shows low sensitivity to an increasing number of devices, and maintains its effectiveness even under limited-data conditions.