Predictive Dynamic Virtual Machine Scaling for Federated Learning Over Edge-Cloud Interworking

The potential of multiaccess edge computing (MEC) to enhance Internet of Things devices with limited resources is examined in this research. We propose a novel predictive virtual machine scaling method that dynamically balances local processing and cloud offloading for federated learning (FL) applications in MEC contexts. Our in-depth analysis of the virtual machine scaling procedure highlights how it affects the computational burden and latency of FL services. Extensive tests on MNIST and CIFAR-10 datasets show the efficiency and flexibility of our method under different computing loads. Our technique performs better than benchmarks in terms of resource utilization, model convergence, and straggler mitigation. Our method maintains 95.54% resource utilization efficiency across a range of workloads while achieving a 20.73% decrease in total latency for FL jobs. Its adaptability in FL settings is shown by its strong performance on the more complicated CIFAR-10 dataset (91.73% accuracy) and the MNIST dataset (95.34% accuracy).