Recently, the top global security conference ACM CCS and the top journal IEEE TDSC have released the list of selected papers. Two privacy computing innovation technologies from Ant Science & Technology have been accepted, marking Ant Science & Technology's continuous leadership in the field of Privacy-Preserving AI technology. These two researches focus on the Gradient Boosting Decision Tree (GBDT) model, which is most widely used in cross-organizational joint modeling. By innovating privacy-preserving algorithms, they solve the technical challenges of achieving high-performance computing while ensuring data privacy in joint modeling and joint inference.

The two research achievements are "Gibbon: Faster Secure Two-party Training of Gradient Boosting Decision Tree" (Gibbon: A Faster 2-Party Secure GBDT Training Framework), accepted by ACM CCS 2025, and "Privacy-preserving Decision Graph Inference from Homomorphic Lookup Table" (Privacy-Preserving Decision Graph Inference Based on Homomorphic Lookup Table), accepted by IEEE TDSC.

GBDT models (including XGBoost, LightGBM, etc.) are decision tree algorithms based on gradient boosting. They are highly interpretable and have fast prediction speed, and are widely used in marketing and risk control scenarios. They are the most popular and commonly used algorithms in cross-organizational joint modeling. However, during multi-party collaborative training and inference, there has long been a dilemma of "strong security leads to low efficiency, and high efficiency brings more risks."

Currently, the industry mostly adopts the Federated Learning (FL) approach. Although it has higher performance, it carries potential information leakage risks. For example, the "Privacy Computing Product General Security Level White Paper" released by the Privacy Computing Consortium in 2024 analyzed and disclosed the information leakage risks in the most popular FL solution, SecureBoost.

Ant Science & Technology has taken a different path, choosing the Multi-Party Computation (MPC) technology route with higher security levels but greater performance challenges. Through deep collaboration between GBDT algorithms and advanced cryptography, they have achieved dual breakthroughs in security and efficiency:

  •  In terms of training: proposed a new secure two-party GBDT training framework called Gibbon, which can increase the training speed by 2 to 4 times compared to the current most advanced MPC scheme "Squirrel" (USENIX Security 2023), and its performance even exceeds the open-source implementation of the federated learning route SecureBoost.

  • l In terms of inference: innovatively proposed homomorphic lookup table technology to achieve privacy-preserving decision graph inference, supporting models such as GBDT, decision trees, and scorecards. Among them, the inference efficiency of GBDT and decision trees is improved by 2 to 3 orders of magnitude.

Currently, the above research achievements have been applied to Ant Science & Technology's privacy computing product series, fully supporting secure, high-performance, and practical data collaboration across institutions.

Ant Science & Technology has built a privacy computing product matrix covering multiple scenarios: including the Trusted Data Circulation Platform FAIR for data infrastructure; the privacy computing solutions Morse for financial and marketing scenarios; the encrypted middleware that provides embedded privacy computing capabilities for AI, BI, and business systems in a lightweight middleware form; and the large model privacy protection products that provide comprehensive data and model privacy protection for large language model applications.

ACM CCS is an internationally recognized flagship conference in the field of information security, listed as a CCF-A class conference by the Chinese Computer Federation (CCF). IEEE TDSC is an authoritative academic journal published by the IEEE Computer Society, focusing on research areas such as trusted computing and secure computing. It is also a CCF-A class journal, representing the highest academic level in this field.