EnFed: An Energy-aware Opportunistic Federated Learning in Resource Constrained Environments for Human Activity Recognition
Anwesha Mukherjee, Rajkumar Buyya
TL;DR
EnFed introduces an energy-aware opportunistic federated learning framework to support personalized human activity recognition when server connectivity or device battery is limited. The method leverages incentive-based handshakes with nearby devices to exchange and aggregate updated models, optimizing for accuracy while minimizing training time and energy on resource-constrained devices. By evaluating LSTM and MLP backbones on two activity datasets, EnFed achieves near state-of-the-art accuracy with substantial reductions in training time and energy compared to CFL and DFL, and significantly faster response times than cloud-only processing. This approach offers practical privacy-preserving, low-latency activity recognition suitable for mobile and edge environments, with clear avenues for future security and mobility enhancements.
Abstract
This paper proposes an energy-efficient federated learning method and its application in human activity monitoring and recognition. In the proposed approach, the device that needs a model for an application requests its nearby devices for collaboration. The nearby devices that accept the request, send their model updates to the requesting device. The device receives the model updates from the collaborators and performs aggregation to build its model. As mobile devices have limited battery life, the number of rounds is decided based on the desired accuracy level and battery level of the requesting device. The performance of the proposed approach is evaluated with respect to prediction accuracy, training time, training energy consumption of the device, and response time. We have used two different datasets for performance evaluation. The first dataset contains different types of physical activities and the respective calorie burn. The second dataset is a human activity recognition dataset that considers six types of physical activities. The experimental results show that using the proposed method the training time and training energy consumption of the device are reduced by approximately 59% and 19% for the first and second datasets respectively, than the decentralized federated learning approach, while using LSTM as the underlying data analysis model. The results also present that the proposed method reduces the training time and energy consumption by approximately 55% and 72% for the first and second datasets respectively, than the decentralized federated learning approach while using MLP as the underlying data analysis model.