Federated Deep Learning in Intelligent Urban Ecosystems: A Comprehensive Review of Privacy-Preserving Advancements and Applications in Smart Cities, Homes, Buildings, and Healthcare Systems

Abstract

Overview

Key Contributions

Methodology

This review adopts a systematic review method to aggregate the state-of-the-art in Federated Deep Learning (FDL) in intelligent urban ecosystems. Literature searches were conducted in IEEE Xplore, ACM Digital Library, SpringerLink, ScienceDirect, and PubMed (for healthcare-related FDL). Search queries combined core terms (federated deep learning OR federated learning) with domain-specific terms (smart city OR smart cities, smart home OR smart building OR smart buildings) and privacy/security terms (differential privacy, secure aggregation, homomorphic encryption, trusted execution environment, blockchain). Boolean operators and phrase searches were used; the time window was 2018–2025, with emphasis on high-impact works from 2019–2025. Inclusion criteria: papers explicitly discussing FL/FDL with privacy or security measures and relevance to urban ecosystems (smart cities, homes/buildings/IoT, or healthcare). Exclusion criteria: non-urban FL applications, work without methodology or experimental validation, and papers without quantitative assessment. This process yielded approximately 174 high-quality references, organized through a PRISMA-style study selection flow and coherent taxonomy of FDL across urban domains.

Results & Findings

The review demonstrates substantial progress in FDL for intelligent urban ecosystems from 2018–2025. Algorithmic advances (FedProx, SCAFFOLD, and variants) improve convergence by 15–20% in heterogeneous urban settings. Hierarchical edge–fog–cloud architectures reduce communication latency by 30–50% in large IoT federations, while energy-efficient methods yield 40–70% energy savings in urban IoT deployments. Domain-specific applications show concrete efficacy: spatio-temporal graph neural networks for traffic prediction on benchmarks like METR-LA; federated CNN/U-Net variants for multi-institutional medical imaging on BraTS; and LSTMs for non-intrusive load monitoring in smart buildings. Hybrid privacy approaches combining differential privacy and secure aggregation provide strong privacy guarantees. Nevertheless, several limitations persist: non-IID data distributions cause client drift and accuracy drops; communication overhead remains high (~100× bandwidth before compression); adversarial attacks can reduce global aggregation accuracy by 30–50%; fairness gaps produce 15–25% performance disparities across demographic groups; and energy consumption on battery-powered devices is 2–5× higher per epoch than centralized training. Pilot deployments in Singapore's traffic monitoring and multi-hospital imaging illustrate viability but highlight the need for standardized testbeds and MLOps frameworks.