What question did this study set out to answer?

This research aims to optimize energy consumption and model performance in federated learning environments by introducing Smart-FLEAT.

July 31, 2026Discover Internet of Things0 citationsOpen Access

Smart FLEAT for energy efficient scheduling and aggregation in federated learning for cloud edge environments

SCSarvani ChalamalasettyJohnson University TBThulasi BikkuAmrita Vishwa Vidyapeetham PMPouria MortezaaghaUniversity of Ottawa

Key Points

This research aims to optimize energy consumption and model performance in federated learning environments by introducing Smart-FLEAT.
Developed Smart-FLEAT framework including DETA-PS, EPGA, and PFTS components.
Evaluated using Environmental Sensor Telemetry dataset (405,184 readings) partitioned non-IID over 50 heterogeneous devices.
Compared performance with conventional FedAvg and recent energy-aware scheduling methods.
Reduced average energy expenditure per training round by 20–35%.
Model accuracy retention remained over 95%.
Increased overall battery preservation ratio to 85% (± 2%).

Abstract

Federated learning (FL) enables collaborative model training across distributed devices while keeping data localized. However, practical deployment in cloud-edge environments remains challenging due to device energy constraints, heterogeneous resource availability, and communication overhead. To address these challenges, this paper proposes Smart-FLEAT, an energy-efficient federated learning framework that introduces three coordinated components: Dynamic Energy Thresholding and Adaptive Participation Scheduling (DETA-PS), Energy-Proportional Gradient Aggregation (EPGA), and Power-Aware Federated Task Sequencing (PFTS). The framework optimizes device participation, aggregation efficiency, and task allocation to improve model performance while reducing energy consumption. Standard privacy-preserving mechanisms, including differential privacy and secure aggregation, are incorporated as compatible system components without modifying the proposed optimization framework. The framework was evaluated using the publicly available Environmental Sensor Telemetry dataset (405,184 readings) from Kaggle, partitioned non-independent and identically distributed (non-IID) (Dirichlet α = 0.4) over 50 heterogeneous devices implemented on Raspberry Pi 4 and NVIDIA Jetson Nano boards. Compared to conventional FedAvg and recent energy- aware baselines such as Energy-Driven Scheduling for Joint Federated Learning (EDS-JFL), Smart Federated Learning for Energy-Aware Task Scheduling (Smart-FLEAT) lowered average energy expenditure per training round by 20–35%, preserved model accuracy retention at over 95%, and raised the overall battery preservation ratio to 85% (± 2%). The results highlight significant gains in energy management, balanced device involvement, and better interoperability in privacy-sensitive cloud-edge federated setups. By extending operational durations and balancing contributions from heterogeneous nodes, Smart-FLEAT supports dependable, long-running IoT applications in cloud ecosystems—particularly environmental monitoring networks and smart farming scenarios—where sustained edge participation is critical.

AI에게 질문

Bookmark

View Full Paper