Federated Learning at the Edge: Enhancing Privacy and Efficiency in Distributed AI Systems
Kofi Sarpong Adu‐Manu, Prince Samuel Kyeremanteng, Prince Ofori, Charles Adjetey, Pachomius Kwaku Lawson, Eugene Cobbah · IntechOpen eBooks · 2025
The growth of data from edge devices has increased the demand for distributed machine learning frameworks that protect privacy while remaining scalable. Federated learning (FL) enables decentralised model training without sharing raw data. This chapter examines FL in edge computing, focusing on its architecture, communication efficiency, privacy, and robustness. We present a hierarchical optimisation framework that includes device profiling, model allocation, and secure aggregation. Gradient compression, over-the-air aggregation, and Byzantine-resilient defences were evaluated for edge contexts. Security methods, including differential privacy and homomorphic encryption, are assessed. Experiments on the FEMNIST and Shakespeare datasets using FedAvg, FedProx, and FedNova showed that FedNova exhibited superior convergence, whereas FedProx demonstrated better dropout resilience. The communication overhead was reduced by 35%, resulting in a 3% loss in accuracy due to compression. Case studies highlight the deployment strategies. We explore the integration of 6G networks and blockchain technology to enhance trust and security. This chapter identifies challenges, including personalisation, incentive-compatible participation, and system optimisation for trustworthy FL across edge environments.