Communication efficient federated learning with data offloading in fog-based IoT environment

Abstract

Federated Learning (FL) has become a popular distributed machine learning technique that preserves privacy of data set generated by the Internet of Things (IoT) devices. However, straggler effect, and communication delay are the two major issues that should be properly taken care during model preparation of any FL process. In this study, we propose a novel and efficient scheme, called Fog Cloud Assisted Federated Learning (FCAFL) in IoT platform that address these issues. Specifically, the proposed scheme is shown to reduce the overall communication delay of the global model preparation in the FL process. It is also able to resolve straggler effect by employing clustering of the IoT devices with an efficient scheme of cluster head (CH) selection based on minimum intra-cluster distance and maximum residual energy. In order to reduce communication delay further, we also propose an efficient scheme of fog node selection using Genetic Algorithm (GA) by formulating the data offloading as a single objective problem with latency as an offloading parameter. There are some existing schemes which address straggler effect and communication delay. However, none of their approaches incorporate explicit clustering technique utilizing residual energy and minimum intra-cluster distance of the CHs to select clients, aiming to address the straggler effect. Additionally, their works do not involve any data offloading as a means to reduce communication delay in the FL process. The simulation results demonstrate better performance of the proposed FCAFL as compared to the traditional FL and existing approaches, namely, Hierarchical Federated Edge Learning (HFEL) and FedFog with respect to communication cost. The FCAFL also outperform traditional FL and HFEL with respect to training and testing accuracy of the produced FL model. The validity of the result is additionally proved through a widely recognized statistical technique, the analysis of variance (ANOVA), followed by a least significant difference (LSD) post-hoc analysis.