Shared Intelligence vs. Federated Learning: Techniques, Challenges, and Ant Group’s Practical Experience

The piece begins by questioning the differences between shared intelligence and federated learning and introduces Ant Group’s shared‑intelligence framework, which combines Trusted Execution Environments (TEE) and Multi‑Party Computation (MPC) to address privacy and usability trade‑offs in AI for finance.

It explains that traditional gradient‑sharing methods used in federated learning are vulnerable to reconstruction attacks, illustrating how attackers can recover original inputs from shared gradients and describing the evolution from DLG to improved attack methods.

To mitigate these risks, Ant Group proposes a layered solution that integrates MPC, TEE, and differential privacy, enhancing data preprocessing (e.g., privacy‑preserving PCA), model training (using semi‑honest servers, Bayesian neural networks, and SGLD), and inference while balancing efficiency, accuracy, and security.

The article details practical deployments, such as joint risk‑control with Jiangsu Bank, privacy‑preserving recommendation systems, and other financial use cases, showing significant performance gains and privacy protection.

Finally, it surveys the four major privacy‑preserving technologies—MPC, TEE, differential privacy, and federated learning—compares their strengths and weaknesses, and outlines future research directions for scaling, improving hardware constraints, and lowering the engineering barrier for broader adoption.