Energy-Efficient Federated Learning for Privacy-Preserving Social Work IoT

Abstract

The incorporation of Artificial Intelligence (AI) and the Internet of Things (IoT) within the domain of social work represents a transformative shift towards proactive, data-driven interventions. Nonetheless, the implementation of such technologies is impeded by the sensitive nature of client information and the imperative for solutions that are both scalable and energy-efficient. This research presents FedSW-Arch, an innovative, privacy-preserving system architecture that leverages Federated Learning to enable model training directly on edge devices, thereby ensuring that raw data remains within a secure and trusted environment. The primary contribution of this study is the introduction of a Communication-Aware Federated Averaging (CA-FedAvg) algorithm, designed to optimize the trade-off between model accuracy and the communication overhead characteristic of Low-Power Wide-Area Networks (LPWANs). Experimental simulations conducted on a Human Activity Recognition (HAR) task demonstrate that FedSW-Arch, when combined with CA-FedAvg, achieves a reduction in data transmission exceeding 60% and a 55% decrease in energy consumption relative to traditional FL approaches, while incurring only a marginal accuracy loss of less than 2%. This work lays the groundwork for a scalable and ethically sound framework for the integration of advanced technologies in social work practice.