Reproduce “Bilateral Differentially Private Vertical Federated Boosted Decision Trees” on OpenResearch

About this paper

Bokang Zhang, Zhikun Zhang, Haodong Jiang, Yang Liu, Lihao Zheng, Yuxiao Zhou, Shuaiting Huang, Junfeng Wu

MaskedXGBoost enables differentially private vertical federated learning for XGBoost, introducing a novel noise mechanism that minimizes utility impact while providing bilateral privacy guarantees for both active and passive parties. The method demonstrates significant improvements in both model utility and computational efficiency over existing privacy-preserving techniques, achieving up to 6.72x faster training times and maintaining higher AUC scores.

Key insights

• MaskedXGBoost introduces carefully calibrated noise mechanisms to achieve differential privacy, avoiding computationally expensive cryptographic techniques.
• The algorithm protects the privacy of both the active (label holder) and passive (feature holder) parties, offering bilateral differential privacy guarantees.
• Noise is introduced into the null space of the arithmetical operations within XGBoost, minimizing its impact on model utility while maintaining strong privacy.

Summary by alphaXiv.