Federated Learning: Background, Techniques, Applications, and the FATE Open‑Source Platform

In this talk, senior researcher Fan Tao from WeBank introduces the background, types, technical implementation, and applications of federated learning, and presents the open‑source industrial platform FATE.

Agenda

Federated learning background

Vertical federated learning

Horizontal federated learning

Application cases

FATE: federated learning open‑source platform

Background

Data privacy and security have become global concerns, and data silos hinder cross‑organization collaboration. Federated learning enables joint modeling without moving raw data, addressing regulatory and privacy constraints.

Challenges in AI deployment

Poor data quality, noisy labels, and fragmented data sources.

Increasingly strict data‑regulation laws in China.

Need for a privacy‑preserving technical ecosystem.

Technical ecosystem

Data isolation – data never leaves the owner.

Lossless performance – federated models match centralized ones.

Peer‑to‑peer collaboration – no single party dominates.

Mutual benefit – both data providers and model users gain value.

Federated learning taxonomy

Vertical federated learning – same users, different features.

Horizontal federated learning – same features, different users.

Federated transfer learning – both users and features differ.

Vertical federated learning

Used for joint credit‑risk modeling where one party holds labels (Y) and the other holds additional features (X). Privacy is protected by RSA‑based ID matching and homomorphic encryption, ensuring only the intersection of user IDs is revealed.

Horizontal federated learning

Applied to anti‑money‑laundering and other scenarios where parties share both features and labels. Techniques include homomorphic encryption, secret sharing, and SecureAggregation to hide model updates.

Application cases

Bank‑regulator joint anti‑money‑laundering modeling.

Internet‑bank joint credit risk modeling.

Insurance personalized pricing.

Small‑business credit risk assessment.

Computer‑vision tasks with federated data.

FATE platform

FATE (Federated AI Technology Enabler) provides industrial‑grade federated learning capabilities, supporting multiple algorithms, secure computation protocols (homomorphic encryption, secret sharing, hash), and a unified workflow consisting of EggRoll (distributed compute/storage), Federated Network (cross‑site communication), FATE‑FederatedML (algorithm library), FATE‑Flow & FATE‑Board (orchestration and visualization), and FATE‑Serving (online inference).

Architecture and workflow

The system follows a pipeline: federated statistics → federated feature engineering → federated model training → federated online inference, with components handling DAG parsing, lifecycle management, experiment tracking, and task scheduling.

Future challenges

One‑stop federated modeling workflow.

Maintainable MPC protocols on WAN.

Secure, auditable cross‑site data transfer.

Adaptation to heterogeneous hardware (CPU, GPU, edge).

For more details, visit the FATE GitHub repository (https://github.com/FederatedAI/FATE) and the official website (https://www.fedai.org.cn/cn/).