zkDaily - 2026-05-12

零知识证明 zkDaily

ZKP Frontier Tracker 🎯

Tue

05.12

2026

Zero-Knowledge Proofs for Gradient Boosted Decision Trees

Paper

https://eprint.iacr.org/2026/907

Jiacheng Gao GBDT KZG ML

Gao et al. proposed Terrae in their paper, a ZK proof system for quantized GBDT training and inference based on KZG commitments, leveraging domain-lifting and interleaving batching to reduce prover time.

Notes

GBDT widely used in finance, healthcare, requiring auditability and privacy.
Existing methods rely on generic ZK compilers or many independent constraints, leading to high prover cost.
Terrae introduces domain-lifting and interleaving batching to reduce commitments.
Novel histogram proof for converting sample data to frequency representation.
Significantly reduces proof generation time with small proof size overhead.
Based on KZG polynomial commitments, suitable for quantized models.

Collected by @icerdesign

零知识证明 zkDaily

Q&A Deep Dive 💬

Tue

05.12

2026

What is GBDT and why do GBDT systems need zero-knowledge proofs?

GBDT (Gradient Boosted Decision Trees) is a machine learning model that combines many decision trees in a boosting process, widely used in finance, healthcare, and recommendation systems. Since models are often trained and served by external providers, users want to verify predictions or training correctness without revealing private data or model parameters, making zero-knowledge proofs essential for privacy and trust.

What is the main optimization introduced by Terrae?

Terrae leverages the structure of GBDT training and introduces domain-lifting and interleaving batching to combine many constraints into a single proof claim.

What is domain-lifting batching?

It is a batching technique for linear constraints that lifts constraints from different domains into a larger shared domain, reducing separate commitments.

Prompted by @icerdesign