Published Paper Details:

Federated Learning allows collaborative model training without sharing raw data, making it ideal for privacy-sensitive Internet of Things (IoT) settings. However, deploying it in real-world scenarios is tough due to differences between devices, limited energy resources, and risks like model poisoning. These challenges create a difficult balance between efficiency, reliability, and model performance. This paper presents a unified orchestration framework for federated learning in a serverless edge-to-cloud environment. The framework includes two main parts. First is an energy-aware client selection mechanism called Adaptive Client Triage (ACT). It assesses devices based on energy levels, computing power, and data relevance to improve training efficiency and speed. Second is a Hybrid Security Aggregation (HSA) protocol. This combines lightweight statistical filtering at the edge with additive homomorphic encryption for safe aggregation in the cloud. This approach boosts protection against harmful updates while keeping overhead low. The system is built using a serverless architecture that spreads computation across edge and cloud resources to enhance speed and reduce costs. Experiments on the FEMNIST dataset show that the framework significantly cuts overall energy use and maintains strong model accuracy, even when facing adversarial attacks. It outperforms traditional federated learning methods.