Benchmarking Federated Learning Applications in Cloud-Edge-IoT Deployments

Federated Learning (FL) enables privacy-preserving model training across resource-constrained IoT devices without sharing raw data. Deploying FL in realistic cloud–edge–IoT infrastructures demands comprehensive benchmarking beyond accuracy to include system-level metrics such as CPU/memory usage, disk I/O, and network traffic. This paper presents FedBench , an extensible platform that automates FL deployment and evaluation in heterogeneous IoT environments. FedBench captures standardized KPIs aligned with the REDRAW framework, supports constrained edge devices via lightweight proxies, and enables custom model/dataset integration through simple Python interfaces. We demonstrate FedBench through two case studies: Fashion-MNIST classification across heterogeneous devices and sleep apnea detection using edge-collected physiological signals. Experimental results show that the federated apnea-detection model achieved average test accuracies of 0.88, 0.87, and 0.86 on heterogeneous clients, while CPU utilization remained below 33% and RAM usage below 50% of device capacity. These results validate the ability of the platform to quantify model performance versus resource trade-offs, supporting optimized FL deployments for IoT applications.