On the Effectiveness of Task Off-loading in Edge based Federated Learning Environment (To Appear)

Abstract

Federated Learning (FL) is a popular approach for distributed machine learning while protecting data privacy, especially in smart city applications. However, FL at the edge faces challenges due to high communication overhead and limited computational resources. Recently, Federated Prompt Learning (FPL) has been developed to reduce data transfer costs by sharing sophisticated prompts instead of model weights, shifting the bottleneck from communication to computation scheduling.
In this paper, we use queueing theory as a lightweight evaluation mechanism to investigate the effectiveness of task offloading in an edge-based FPL environment. We analyze the impact of scheduling (both preemptive and non-preemptive) based on the priority of inference and training tasks on task execution latency. To address the limitations of these conventional methods, we propose a novel offloading strategy that distributes inference and training tasks across multiple edge servers. Experimental results demonstrate that the proposed method significantly reduces inference latency while maintaining efficient training and outperforms the baseline scheduling method. The paper provides insights into load balancing optimizations in edge-based federated learning environments, contributing to more responsive, scalable, and autonomous systems.