Vertical Federated Learning in Finance 金融领域的纵向联邦学习

Federated learning can be categorized into horizontal, vertical, and federated transfer learning based on how participants’ data is distributed.

Horizontal FL splits by samples and allows relatively general solutions. Vertical FL is more challenging and widely applied in industry — different parties hold different feature dimensions of the same samples (due to different business functions).

In risk control scenarios, vertical FL is particularly prevalent: labels are held by only one party (Guest), while others (Host) hold partial features. The Guest wants to improve model performance through collaboration while both parties must ensure data security.

SecureBoost

In risk control, tree models (especially XGBoost) are widely used for their interpretability. SecureBoost extends XGBoost to vertical FL using homomorphic encryption:

1. Encrypted sample alignment using Private Set Intersection (PSI)
2. Encrypted model training: Guest computes first/second-order gradients, encrypts and sends to Host parties. Host computes aggregated split gains under encryption and returns to Guest for decryption. Best splits are selected without exposing raw data.

Interpretability Requirements

In single-party sub-model scenarios (Guest has only labels, Host has features), new requirements emerge:

• Guest needs to understand node feature meanings for model interpretability
• But knowing predictions + tree structure could allow Guest to reverse-engineer Host’s feature values
• Conversely, Host cannot access Guest’s labels or leaf weights

Current SecureBoost cannot simultaneously satisfy both requirements — a key limitation for real-world deployment.

Efficiency Challenges

Homomorphic encryption requires massive ciphertext communication per training round. SecureBoost provides high security but at enormous communication cost, significantly reducing practical usability — a major obstacle for deployment.