TinyFL_HKD: Enhancing Edge AI Federated Learning With Hierarchical Knowledge Distillation Framework

Abstract

With the rapid evolution of artificial intelligence (AI) and the Internet of Things (IoT), machine learning is increasingly being integrated into embedded systems, bringing computational capabilities closer to where data are generated. This article introduces a tiny federated learning framework, which concerns privacy, personalized training, and the constrained computational resources of edge platforms by introducing a novel hierarchical knowledge distillation (HKD), called TinyFL_HKD. The HKD leverages hierarchical learning and advanced encryption security (AES) schemes to ensure data privacy and security. It employs knowledge distillation to reduce model complexity for implementation in edge devices and enhance personalization. The performance of TinyFL_HKD is introduced by using two datasets: the tool wear dataset and the PHM 2010 Data Challenge dataset. Experimental results indicate that the HKD framework surpasses traditional federated averaging (FedAvg) and personalized federated learning (PFL) algorithms in both model accuracy and computational efficiency. This establishes HKD as a resilient solution for edge AI applications.