Real-Time Prediction Using Fog-Based Federated Learning and Genetic Hyperparameter Optimisation

Federated Learning (FL) has empowered advancements in machine learning by using model sharing as an alternative to data sharing. This feature avoids uploading huge amounts of continuous data streams onto a central server thus reducing the need for high storage capacity of a central server. However, classical FL relies on a single central server for global model aggregation, which causes resource bottlenecks and poor client performance. Moreover, FL often deals with heterogeneous and non-IID data, which requires careful hyperparameter tuning. To overcome these issues, we propose a novel FL framework that leverages fog nodes and the Genetic Algorithm (GA) to address the challenges of single-point failure, communication efficiency, and client-side hyperparameter optimization. We use fog nodes as intermediate aggregators between the clients and the central server and apply the GA at the fog level to optimize the hyperparameters for each client. Fog nodes reduce the communication overhead and increase the robustness of the FL framework, while GA exploits the exploration and exploitation trade-off to find the optimal hyperparameters. Empirical results, obtained on real-time datasets demonstrate that adopting the proposed framework significantly enhances the performance of state-of-the-art versions of the FL framework.