Practical Experience of Federated Learning in Urban Applications

Federated learning and privacy computing are introduced as technologies that enable valuable data exchange without exposing raw data, with three main branches: trusted execution environments, secure multi‑party computation, and federated learning itself.

The article describes the federated learning architecture, including functional layers such as application, service, task training, algorithm, data preprocessing, data sources, and runtime environments, and outlines supporting components like feature engineering, data services, cross‑domain applications, and model lifecycle management.

A detailed technical stack is presented, covering application scenarios (credit, finance, marketing, public security), data services (feature, model, profile, label, metric libraries), cross‑domain capabilities, security encryption (OAuth2, RSA, AES, financial‑grade encryption, HTTPS, IP whitelisting), hardware security (SGX), and data storage/computation protections (encryption, homomorphic encryption, differential privacy, secure multi‑party computation).

The data flow of federated learning is illustrated through a step‑by‑step process involving data import, preprocessing, gradient saving, encrypted transmission via SDK, temporary encrypted storage, arbitration node handling, and final encrypted return to participating parties.

Several city‑level case studies are highlighted, including credit scoring, risk management, intelligent site selection, and a city‑wide credit system, demonstrating how federated learning breaks data silos to improve model accuracy and decision making.

The current status discusses standardization efforts (IEEE federated learning standards), platform industrialization (federated digital gateway), and ongoing research in secure multi‑party computation, while the future outlook emphasizes expanding data resources, breaking institutional data boundaries, industry collaboration, and broader adoption of federated learning across sectors.