Federated Learning-based Adaptive Idle-Time Training for Energy-Efficient Edge Nodes

Document Type

Conference Proceeding

Source of Publication

Digest of Technical Papers IEEE International Conference on Consumer Electronics

Publication Date

2-3-2026

Abstract

Federated Learning (FL) enables collaborative model training across distributed edge devices without sharing raw data, yet its deployment in energy-constrained IoT environments remains inefficient. This paper proposes Fed-Idle, an Adaptive Idle-Time Training framework that improves energy efficiency by aligning local computation with each device's power and workload context. Each node monitors its battery, CPU utilization, and charging state, activating training only during idle or charging periods. A centralized coordinator aggregates updates from active nodes using staleness-aware weighting. Experiments on the CIFAR-10 dataset show that Fed-Idle achieves up to 40% lower energy consumption, 25% faster convergence, and a 5% reduction in client dropout, while maintaining comparable accuracy to FedAvg and FedProx. These results demonstrate that idle-time-aware scheduling provides an effective foundation for sustainable federated edge intelligence.

ISBN

[9798331553432]

ISSN

0747-668X

Publisher

IEEE

Disciplines

Computer Sciences

Keywords

Adaptive scheduling, Edge computing, Energy efficiency, Federated learning, Internet of Things (IoT)

Scopus ID

105037341455

Indexed in Scopus

yes

Open Access

no

Share

COinS