Abstract
Federated learning (FL) has emerged as a promising paradigm for collaborative machine learning across decentralized edge devices while preserving data privacy. However, deploying FL in heterogeneous embedded networks—characterized by varying computational capacities, communication constraints, and data distributions—poses significant challenges. This article reviews and extends existing FL architectures tailored for such environments, focusing on privacy preservation. We propose a taxonomy of FL frameworks including synchronous, asynchronous, vertical, and transfer learning approaches, and evaluate their suitability for embedded systems. Through simulation experiments on representative embedded network scenarios, we compare communication efficiency, convergence speed, and accuracy under differential privacy and secure aggregation. Results show that asynchronous FL with selective participant aggregation achieves 15% higher throughput than synchronous FL in high-heterogeneity settings. Additionally, a hybrid blockchain-based FL architecture enhances auditability with moderate overhead. Our findings highlight the trade-offs between privacy, efficiency, and model accuracy, and provide design guidelines for deploying privacy-preserving collaborative AI on resource-constrained devices. This work underscores the need for adaptive aggregation strategies and lightweight cryptographic primitives for embedded FL.