Shared Intelligence: Combining Trusted Execution Environments and Multi‑Party Computation for Privacy‑Preserving Machine Learning at Ant Group

With the rise of artificial intelligence, data silos and privacy protection have become hot topics. Ant Group proposes a shared‑intelligence solution based on two key technologies: Trusted Execution Environments (TEE) and Multi‑Party Computation (MPC). The paper details the technical specifics, compares different approaches, and describes real‑world applications within Ant Group.

Overview – Data is the core asset in the internet era; high‑quality, large‑scale data improves machine‑learning models, but sharing data raises privacy leakage and misuse concerns. GDPR and strict financial regulations intensify the need for secure data collaboration.

Industry Progress – Current research includes privacy‑preserving machine learning, federated learning, split learning, and secure computation, each with distinct trade‑offs in security, communication, computation, and model accuracy.

Necessity of Shared Intelligence – Ant Group needs a solution that simultaneously satisfies high data‑security requirements and massive data volume for financial services. Existing single‑technology approaches cannot meet all business scenarios.

Technical Evolution

1. Privacy‑Preserving Data Transform (PPDT) : Data is transformed into a low‑dimensional, irreversible representation with differential‑privacy noise, enabling joint modeling while protecting raw values.

2. TEE‑Based Shared Intelligence : Uses Intel SGX enclaves and remote attestation to create a trusted third‑party execution environment, eliminating the need for a privileged administrator. A large‑scale SGX cluster extends compute capacity for massive data workloads.

3. MPC‑Based Shared Intelligence : Supports vertical and horizontal data partitioning. Implements cryptographic primitives (secret sharing, homomorphic encryption, garbled circuits), secure operators (matrix multiplication, comparison), a DSL for algorithm development, and a compiler that selects optimal secure operators.

Development Trend – TEE offers low‑cost centralized training, while MPC provides transparent, strong security for inference. Combining both (TEE for training, MPC for prediction) yields a practical, high‑performance solution.

Applications at Ant Group

Joint credit scoring with Zhonghe Rural Credit Union using TEE‑based shared intelligence, serving millions of rural customers.

RiskGo joint risk control (BESA) leveraging TEE to protect cross‑enterprise payment data, improving detection accuracy and reducing loss.

Joint lending with Jiangsu Bank via MPC, increasing KS metric by over 50% and lowering risk.

Challenges and Outlook – Building industrial‑grade, privacy‑preserving ML models remains a major challenge. Ant Group will continue to evolve its multi‑technology stack, integrate TEE and MPC, and contribute to AI research and standards.

References

[1] Mohassel & Zhang, SecureML, 2017. [2] Kairouz et al., Federated Learning, 2019. [3] Vepakomma et al., Split Learning for Health, 2018. [4] Arnautov et al., SCONE, 2016. [5] Zhu et al., Deep Leakage from Gradients, 2019. [6] Kantarcioglu & Clifton, Privacy‑Preserving Distributed Mining, 2004.