Baidu Data Federation Platform: Architecture, Applications, Federated Learning, and Explainability

The talk, presented by Baidu senior researcher Dr. Liu Ji, introduces the Baidu Data Federation Platform and its research applications.

Architecture Overview : The platform is organized into four layers – Data Layer (maps, search, city, POI data), Platform Layer (provides compute and data processing), Module Layer (epidemic analysis, federated learning/computation, explainability, AutoDL), and Application Layer (supports collaborations with academia, research institutes, and government).

Technical Capability Framework : Describes the four data types and the four module functions, emphasizing privacy‑preserving joint modeling for users with partial or confidential data.

Platform Deployment : Supports two user categories – those with private server clusters (e.g., finance, hospitals) that connect securely to Baidu's servers, and those without private infrastructure that upload data to Baidu Cloud, where dynamic server clusters are created for secure joint modeling.

Collaborative Research Examples :

Epidemic prediction using the SIR‑X model combined with Monte‑Carlo methods and Baidu migration data, accelerated via parallel cloud computing, achieving accurate forecasts and revealing correlations with GDP and search frequencies.

Early‑stage epidemic studies leveraging satellite imagery, search indices for symptoms, and population flow analysis to identify the onset of community response.

Shared‑vehicle optimization using Baidu map data, formulating a cost model and greedy algorithms to dynamically schedule routes, reducing dispatch time by up to 82% and costs by 47‑70%.

Federated Learning :

Three types – horizontal (different users with same features), vertical (same users with different features), and hybrid.

Horizontal federated learning with PaddleFL uses a parameter‑server architecture; dynamic sampling balances load across heterogeneous compute resources, shortening training time without sacrificing accuracy.

Vertical federated learning introduces Secure‑GBM, employing privacy‑preserving set intersection (OT‑PSI), homomorphic encryption for gradient exchange, and decentralized model updates to ensure data confidentiality.

Explainability :

Highlights the need to interpret black‑box deep models; demonstrates using LIME with tree models to assign meaningful weights to features such as credit score, income, and loan history in financial risk prediction.

Conclusion : The article summarizes the platform’s architecture, showcases its applications in epidemic modeling and shared‑vehicle research, and discusses federated learning techniques and explainability methods that enable secure, collaborative AI across diverse data sources.