Building and Evaluating Federated Models for Edge Computing

Abstract

Today’s state-of-the-art machine learning (ML) techniques, such as deep learning (DL) networks are typically trained using cloud platforms, leveraging elastic scalability of the cloud. For such processing, data from various sources need to be transferred to a cloud server. While this works well for some application domains, it is not suitable for all applications due to concerns about latency, connectivity, and privacy. For example, sharing life logging photos and videos from cellphones and wearable devices can cause privacy concerns for users, and transferring the unstructured data can burden the communication network. With the increase of such applications, federated learning (FL) is proposed as a distributed ML solution for learning on edge devices, such as cellphones and wearable devices. In FL, clients collaboratively train a model on their device without sharing their data. Each client trains a local model with their data and shares the model parameters with a FL server to aggregate and build a global model. Shifting from traditional ML techniques to federated solutions requires comparing these two approaches. Moreover, users need to study the performance of FL models to decide if federation is feasible for their learning task. In this paper, we propose an automated solution to compare centrally trained DL models with federated solutions. The tool allows users to easily analyze the accuracy of federated models for their learning task and study the effect of the federated parameters. We show the features of our tool building central and federated DL models from an input model structure for recognizing images in the MNIST benchmark dataset.