An Efficient and Privacy-Preserving Framework for Edge AI Using Federated Learning and Homomorphic Encryption

Abstract

Edge Artificial Intelligence (Edge AI) supports intelligent processing close to where data is generated. However, training machine-learning models across edge devices is difficult when devices have limited memory, processing power, battery capacity, and network bandwidth. Federated learning (FL) addresses one important concern by allowing devices to train a shared model without sending their raw records to a central repository. Nevertheless, model updates may still expose information, while repeated communication and encryption can increase training time and resource use. This paper reviews an integrated architecture that combines federated learning, homomorphic encryption (HE), client clustering, update pruning, adaptive quantization, and convergence-aware stopping. In the architecture, each client trains locally, removes low-impact update values, compresses the remaining update, and encrypts it before transmission. The server aggregates encrypted updates and coordinates the next training round. The source study reports that its FLHC-QPR approach improves accuracy and reduces latency and memory overhead compared with selected comparison methods on a Human Activity Recognition task. The paper explains the architecture, workflow, potential applications, evaluation measures, limitations, and future research directions. The discussion is based on the supplied research article; performance values are reported from that study and are not presented as newly measured results.