Shared Intelligence vs. Federated Learning: Ant Group’s Privacy‑Preserving Machine Learning Solutions for Finance

Artificial intelligence faces a fundamental dilemma: privacy and usability cannot be fully achieved together, especially under tightening data‑security regulations that have fragmented data into isolated silos. To enable data sharing while protecting privacy, Ant Group (蚂蚁金服) has developed a solution called “shared intelligence” (共享智能), which blends Trusted Execution Environment (TEE) technology and Multi‑Party Computation (MPC).

The article first distinguishes shared intelligence from federated learning. Federated learning, introduced by Google, focuses on cloud‑plus‑edge scenarios (To‑C) with a central server coordinating model updates, whereas shared intelligence targets both To‑B and To‑C contexts, supporting centralized, decentralized, and TEE‑based modes.

Two main technical routes for privacy‑preserving data sharing are highlighted: hardware‑based trusted computing (TEE, e.g., Intel SGX) and cryptographic MPC. TEE enables secure enclaves where data can be processed without leaving the hardware boundary, while MPC allows multiple parties to jointly compute functions without revealing raw inputs.

Ant’s TEE‑based shared learning leverages Intel SGX (and other TEEs) to provide encrypted data‑out‑of‑domain processing, supporting both online prediction and offline training. A custom distributed service framework registers each enclave with a Cluster Manager, synchronizes keys via secure channels, and handles fault tolerance and scaling—challenges that traditional clustering cannot meet on SGX.

For MPC‑based shared learning, Ant builds a three‑layer stack: a security‑technology layer (secret sharing, homomorphic encryption, differential privacy, etc.), a basic‑operator layer (secure matrix operations, activation functions), and a secure‑ML‑algorithm layer (privacy‑preserving LR, GBDT, DNN). The training workflow involves downloading encrypted tools, uploading encrypted data to cloud storage, and orchestrating tasks through a coordinator, task flow manager, and worker nodes that execute secure operators.

Practical deployments in the financial sector demonstrate significant benefits: shared intelligence improves risk control, reduces loan approval time to minutes, and has driven billions of yuan in lending across thousands of villages. The interview emphasizes that financial services, with strict data‑governance requirements, are the primary early adopters of these technologies.

Looking forward, Ant aims to lower adoption barriers by open‑sourcing components, contributing to industry standards, and fostering an ecosystem where enterprises can collaboratively build privacy‑preserving data‑sharing networks, ultimately enabling inclusive finance while safeguarding user privacy.