Design and Implementation of Kubernetes enabled Federated Learning Platform

Abstract

Federated learning (FL) is one of most promising distributed machine learning frameworks to strengthen data privacy and security by making. Specifically, in the FL, the multiple clients are involved to train their deep learning models (i.e., local training) and then aggregate them as a global model at the server. However, due to the heterogeneous nature of multiple distributed clients, it is challenging to implement and deploy the FL framework over various clients. By adopting the cloud native technologies such as containers, and Kubernetes, we design and implement the Kubernetes enabled Federated Learning Platform (KubeFL). The KubeFL helps the deployment of the FL over multiple clients via Docker containers and Kubernetes. To do this, in the proposed KubeFL, each FL framework implemented by the PyTorch is deployed on the Docker container in the Pod at each client (i.e., Node in the Kubernetes) while the one is also conducted at the server (i.e., Master in the Kubernetes). We evaluate its performance through extensive measurement-based analysis on the commercial NVIDIA Jetson TX2 edge devices under various practical configurations.