Privacy Computing: The Federated Learning Three‑Part FIRM Architecture and Its Industrial Applications

Introduction In the era of artificial intelligence, the value of data is increasingly unlocked while user privacy and security management become stricter, making privacy‑preserving computation essential. Privacy computing enables multiple parties to jointly compute without exposing raw data, breaking data silos.

1. Privacy Computing Background The industry sees three main technical paths: federated learning, secure multi‑party computation, and trusted computing, with blockchain often complementing these approaches. Driven by big‑data integration and privacy demands, privacy computing is becoming a fundamental solution for data collaboration across finance, government, healthcare, and other domains.

2. Federated Three‑Part – FIRM Reference Architecture

2.1 FIRM Theoretical Foundations FIRM (open Federated system Interconnection Reference Model) defines a four‑layer stack—communication, data exchange, algorithm, and application—standardizing protocols to enable interoperable federated learning systems.

2.2 Federated Three‑Part: Protocol, Algorithm, Application The federated protocol (communication + data‑exchange layers) standardizes data‑exchange order and encryption methods. Federated algorithms implement multi‑party training for models such as logistic regression and decision trees. Federated applications package these algorithms into business‑level workflows, allowing rapid deployment in vertical scenarios.

2.3 Federated Protocol Details The protocol includes application‑level agreements and public cryptographic components (e.g., AES, SM4, homomorphic encryption, secret sharing) to ensure secure data exchange.

3. Key Technologies of the ZhiBang Platform

3.1 Communication Framework – Ionic Bond Ionic Bond is a lightweight, high‑throughput federated communication framework built on gRPC and Kubernetes, achieving 2‑6× speedup over open‑source alternatives and supporting GB‑level data transfer over both WAN and LAN.

3.2 HeteroDeepFM Scheme To bring DeepFM into federated learning, the HeteroDeepFM solution leverages the FLEX protocol for encrypted data exchange while performing non‑interactive operations (e.g., feature encoding, embedding) locally, enabling secure, lossless deep‑learning model training.

4. Industrial‑Grade Product – ZhiBang Platform

4.1 Marketing Case Using federated learning, a bank and an insurance company jointly identified potential insurance customers, improving KS value by >5% and achieving a >3× increase in conversion rate compared with single‑party models.

4.2 Risk‑Control Case A consumer‑finance company adopted a federated scoring‑card model, raising KS value by >30% and handling ~500,000 monthly inference requests with real‑time latency.

4.3 Government Case The platform enabled secure, cross‑department data analysis for a city, improving administrative efficiency by >2% through real‑time monitoring dashboards.

5. Summary & Outlook Privacy computing is still in its early production stage, but with advancing standards and products it will increasingly support AI applications while safeguarding user data, paving the way for a commercial AI ecosystem.

6. Q&A Highlights • The federated learning framework supports Chinese national cryptographic standards SM2/SM3/SM4. • Public components such as AES, ECDH, homomorphic encryption, and secret sharing are used in different protocol phases. • Incentive mechanisms are based on Shapley value contribution assessment. • The platform supports various association algorithms (e.g., KNN) and real‑time inference with optimized model compression.