Reproduce “Tree-based Models for Vertical Federated Learning: A Survey” on OpenResearch

About this paper

Bingchen Qian, Yuexiang Xie, Yaliang Li, Bolin Ding, Jingren Zhou

This survey paper from Alibaba Group provides a structured overview of tree-based models (TBMs) in Vertical Federated Learning (VFL), categorizing them into feature-gathering and label-scattering approaches. It analyzes privacy protection mechanisms, reviews implementation infrastructures, and empirically demonstrates the trade-offs between model utility, privacy strength, and resource costs, finding comparable performance across TBM categories but higher communication for label-scattering, and inverse relationships between utility/privacy and privacy/cost.

Key insights

• VFL TBMs can be systematically categorized by their communication patterns (feature-gathering vs. label-scattering), each with distinct implications for efficiency and privacy.
• There are fundamental trade-offs in VFL TBMs: increasing privacy protection strength often reduces model utility or significantly increases computational and communication costs.
• Designing robust VFL platforms requires adhering to principles of Flexibility, Extensibility, Scalability, and Security to support diverse TBMs and privacy requirements effectively.

Summary by alphaXiv.