A Kubernetes-Powered Personalized Federated Learning Platform for Resource-Constrained Internet of Medical Things

Federated Learning (FL) is a distributed machine learning scheme that trains a global model across multiple end devices while protecting user privacy by keeping data locally, which has been shown to be useful for smart healthcare. By integrating virtualized container technology into edge platforms, flexible computing resources can be made available to Internet of Medical Things (IoMT) devices, enabling their participation in FL. Some studies have utilized the advantages of Kubernetes (K8s) for fast container deployment, offering general FL platforms for developers to rapidly deploy models. However, there are currently no FL development platforms suitable for smart healthcare, which exhibits unique challenges of personalized medical data and large-sized models for resource-limited IoMT devices. In this paper, we propose a Kubernetes-powered personalized FL (PFL) platform for resource-constrained IoMT. This platform, which incorporates personalized strategies to capture individual features and model compression mechanisms to reduce model size, allows model developers to formulate PFL training tasks for healthcare users to acquire personalized and compressed models, deployable on their devices. Via establishing a real testbed and deploying a demonstrating training task, our experiments verify the platform’s ability to support the development of personalized and compressed models.