An Efficient Architecture for Edge AI Federated Learning With Homomorphic Encryption

Abstract

With the rapid growth of edge AI applications, there is an increasing demand for federated learning (FL) frameworks that are both efficient and privacy-preserving. This work introduces a robust approach that leverages homomorphic encryption (HE) to ensure data confidentiality during decentralized training. To tackle the typical challenges of FL—such as high communication overhead, resource limitations, and convergence inefficiencies—the method integrates dynamic client clustering, quantization-aware training, and structured model pruning. These optimizations collectively reduce latency and memory consumption while accelerating model convergence. Evaluations using the Human Activity Recognition dataset show that the proposed approach outperforms several state-of-the-art FL methods, achieving an average +8.4% improvement in accuracy, -16.2% lower latency, -35.1% reduction in memory usage, and -2.7% lower security overhead. These results demonstrate its suitability for real-time, resource-constrained scenarios in domains like healthcare, IoT, and finance, where maintaining a strong balance between efficiency and privacy is essential.