A Context-Aware Localized Federated Learning Approach for Cloudlet Federation

Abstract

Cloud computing (CC) provides storage and computation capability to perform resource intensive Machine Learning (ML) tasks for prediction and decision making. However, cloud services can introduce latency and bandwidth limitations while transmitting data to a remote cloud. Cloudlets are deployed as stand-alone devices at the network edge to bring computation in the closer proximity of users to address cloud computing challenges. Moreover, cloudlets have insufficient resources to train these resource-intensive deep learning models. Therefore, deployment of cloudlet federation can resolve latency, storage, computation, and bandwidth limitations by offloading tasks to a cloudlet within the federation. The selection of a deep learning model to reduce communication and computation cost is a challenge, as cloudlets are not context-aware in terms of network load and latency in the cloudlet federation. Moreover, deep learning model accuracy and prediction results can be affected if end-user devices are unreliable and provide incorrect data for training deep learning models at cloudlets. A cloudlet federation based novel solution is discussed in this thesis for Federated Learning (FL) that monitors network load and resources using a broker. The broker is a centralized entity that will reside within the federation and the global model will be stored inside the broker to make localized decision. COVID-19 X-ray images are used to train the model. The data is divided into 70% training and 30% validation data sets. Besides cloudlets broker will have its dataset to check the validation accuracy of the global model. The sampling data and network parameters such as available storage will be used to extract a context-aware local model for each cloudlet from the generic model. The local model will be converged using Root Mean Square Proportional (RMSProp) based on the mean squared error loss function as it provides higher training accuracy, irrespective of training data size. Convolutional Neural Network (CNN) is used in this thesis with 6 layers for classification of data, one average pooling layer, a dense and a dropout layer. The trained model on each cloudlet is sent to the broker for aggregation. Two aggregation methods are proposed in the thesis. The first one uses the layer aggregation method and the second one is based on the best model selection. Vertical Federated Learning (VFL) is used for model training. Experiments with method one gives a model accuracy of 86% and loss is 14%. However, the best model selection method generates a model x accuracy of 96% and a loss of 15%. Total time taken for model convergence using conventional FL architecture is approximately 30 minutes. However, the designed architecture takes approximately 28 minutes for completion of global epochs. Based on the results obtained it can be concluded that the designed testbed can identify COVID-19 infected person using X-ray image or CT-Scan with 86% accuracy using FedAvg algorithm and with 96% accuracy using best model selection