Research map / Learning theory and algorithms / Prediction and decision losses
Robust Watermarking on Gradient Boosting Decision Trees
Jun Woo Chung, Yingjie Lao, Weijie Zhao
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| choose watermark samples with Cluster Center Flip instead of random | GBDT watermarking · Dcand=Dtrain avg, ratio 0.01 | watermark effectiveness: 0.819 -> 0.980 | Improved |
| choose watermark samples with Confidence Flip instead of random | GBDT watermarking · Dcand!=Dtrain avg, ratio 0.1 | watermark effectiveness: 0.800 -> 0.947 | Improved |
| choose watermark samples with Cluster Center Flip instead of random | GBDT watermarking · Dcand!=Dtrain avg, ratio 0.1 | adjusted model accuracy: 0.714 -> 0.851 | Improved |
| choose Cluster Center Flip instead of random for robustness | GBDT watermarking · Dcand!=Dtrain avg, ratio 0.1 | fine-tuning robustness: 0.684 -> 0.918 | Improved |
| choose Confidence Flip (max-distance) instead of random for robustness | GBDT watermarking · Dcand=Dtrain avg, ratio 0.1 | fine-tuning robustness: 0.934 -> 0.989 | Improved |
| use Outlier Flip at high watermark ratio | GBDT watermarking · Dcand!=Dtrain avg, ratio 0.1 | watermark effectiveness: 0.863 vs 0.947 for Confidence Flip | Hurt |
| use Outlier Flip at low watermark ratio | GBDT watermarking · Dcand=Dtrain avg, ratio 0.001 | watermark effectiveness: 0.896 vs 0.694 random | Improved |
| use Wrong Prediction Flip when Dcand=Dtrain | GBDT watermarking · Dcand=Dtrain | applicability | Hurt |
| add in-place fine-tuning instead of adding trees | GBDT watermarking · design | robustness to tree pruning | Improved |
| use maximum-distance selection instead of lowest-confidence selection (ablation) | GBDT watermarking · Dcand=Dtrain avg, ratio 0.1 | watermark effectiveness: 0.999 vs 0.999 (Confidence Flip) | No clear effect |
| duplicate watermark samples by factor 5 when Dcand=Dtrain | GBDT watermarking · Dcand=Dtrain | watermark effectiveness | Improved |
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.85)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.84)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.84)
- Decoder Gradient Shield: Provable and High-Fidelity Prevention of Gradient-Based Box-Free Watermark Removal (CVPR 2025 · similarity 0.84)
- Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization (NeurIPS 2025 · similarity 0.84)
- An Undetectable Watermark for Generative Image Models (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
