How Federated Learning Is Breaking Data Silos Across Clouds

Artificial intelligence has evolved from early Dartmouth workshops to breakthroughs like Deep Blue and AlphaGo, and now federated learning emerges as a new technology aimed at solving the data‑island problem in the AI and big‑data era.

With data protection regulations tightening worldwide and moving toward GDPR‑style controls, cross‑cloud data collaboration has become increasingly difficult, prompting the need for secure, privacy‑preserving methods.

Data security is the core premise of large‑scale data cooperation, explains Chen Tianjian, deputy general manager of WeBank’s AI department. He outlines the historical milestones: the 1990s concept of federated databases for storage security, 2010‑2015 secure multiparty computation for computation security, and the 2017 emergence of federated learning to ensure information security.

Federated learning can be categorized into three types—horizontal, vertical, and federated transfer learning. Among them, vertical federated learning has the most business scenarios, and currently only WeBank, Tencent, and VMware have mature vertical solutions.

WeBank has open‑sourced an industrial‑grade federated learning system called FATE under the Linux Foundation. FATE provides a complete, auditable codebase, strong reliability, data‑security guarantees, scalable architecture, and interoperability with other software.

Tencent Cloud’s AI team, led by Lei Xiaoping, views federated learning as a productizable capability. Traditional approaches—direct data sharing after anonymization, separate model training followed by model fusion, and centralized trusted environments—each suffer from compliance, performance, or fairness issues.

Lei proposes three federated learning strategies:

ID matching without exposing query IDs or non‑overlapping IDs.

Secure arithmetic operations that hide each party’s numeric values.

Model transformation that abstracts data exchange into secure arithmetic, reducing interaction.

He notes that federated learning does not solve every data‑fusion security challenge; it mainly addresses typical machine‑learning scenarios, while secure SQL computation remains outside its scope.

To serve B2B customers, Tencent Cloud’s ShenDun federated learning team offers a five‑step workflow—task creation, secure intersection, feature engineering, feature selection, and result visualization—covering use cases such as new‑user acquisition models and homepage prediction models. The underlying training framework builds on WeBank’s mature FATE platform, adding layers for environment security, productization, and scenario‑specific adaptation.

VMware chief architect Zhang Haining, who moderated the event, emphasizes that federated learning is a rapidly growing AI subfield capable of joint modeling while preserving data privacy, turning data into valuable assets and potentially becoming a key driver of future AI innovation.