Communication-Efficient Federated Learning for IoT Edge Computing with LION-Based Optimization

Training a shared model on Internet of Things (IoT) data presents two practical difficulties: devices have limited computing resources, and repeated exchanges with a remote server consume time and bandwidth. Federated learning keeps the training data on the devices, but it does not remove these costs. We propose HELA-FL, a hierarchical framework that places edge servers between IoT clients and a central server. Clients learn from their local datasets, edge servers coordinate intermediate updates and aggregation, and the central server combines the resulting edge models. The framework uses the LION optimizer to support model optimization within this hierarchy. Its purpose is to make use of the greater computing capacity available at the edge while reducing the frequency of global communication. We implemented the framework in Python and evaluated it on MNIST with a non-independent and identically distributed (non-IID) allocation of data. The simulation used five edge servers, each serving twenty clients. We compared HELA-FL with FedAvg, FedBuff, AdamW, FedCMFL, FedEdge, FedNova, FedScalar, and FedProx over twenty training rounds. The reported final accuracy was approximately 94%, with latency of about 12 to 12.5 seconds and a combined cost of approximately 220 units under the study's cost definition. Some baselines reached higher accuracy. The value of the proposed approach therefore lies in the balance between accuracy, latency, and cost in this experiment. These results support further investigation of hierarchical aggregation with LION for IoT edge systems; evaluation on a physical testbed remains future work.