Research map / Adversarial robustness and security / Attacks on text and graphs
Neural Honeytrace: Plug&Play Watermarking Framework against Model Extraction Attacks
Yixiao Xu, Binxing Fang, Rui Wang, Yinghai Zhou, Yuan Liu, Mohan Li, Zhihong Tian
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| use composite trigger instead of white pixel block (ablation) | Neural Honeytrace · CIFAR-10 | average watermark success rate: 37.86% -> 58.60% | Improved |
| use composite trigger instead of semantic object (ablation) | Neural Honeytrace · CIFAR-10 | average watermark success rate: 25.08% -> 58.60% | Improved |
| attacker queries with in-distribution CIFAR-10 instead of TinyImageNet (ablation) | Neural Honeytrace · CIFAR-10 | average watermark success rate: 58.60% -> 29.18% | Hurt |
| attacker queries with CIFAR-100 instead of TinyImageNet (ablation) | Neural Honeytrace · CIFAR-10 | average watermark success rate: 58.60% -> 48.78% | Hurt |
| VGG-16 target with ResNet-18 surrogate (architecture mismatch) (ablation) | Neural Honeytrace · CIFAR-100 teacher-student pairs | minimum watermark success rate: 43.80% | No clear effect |
| increase attacker query sample size from 5,000 to 50,000 (ablation) | Neural Honeytrace · CIFAR-10 | watermark success rate | No clear effect |
| larger d and smaller alpha, beta (ablation) | Neural Honeytrace · CIFAR-10 | watermark strength vs protected accuracy | Mixed |
| increase CLP pruning strength against watermark (ablation) | Neural Honeytrace · CIFAR-10 | watermark success rate | No clear effect |
| replace MEA-Defender with Neural Honeytrace | MEA-Defender · CIFAR-10 | minimum watermark success rate: 11.60% -> 36.60% | Improved |
| replace MEA-Defender with Neural Honeytrace | MEA-Defender · CUB-200 | average watermark success rate: 53.52% -> 44.12% | Hurt |
| replace MEA-Defender with Neural Honeytrace | MEA-Defender · CUB-200 ownership assertion | queries to assert ownership: 24{,}156 -> 590 | Improved |
| replace MEA-Defender with Neural Honeytrace | MEA-Defender · CIFAR-100 | average watermark success rate: 19.80% -> 55.62% | Improved |
Closest papers by meaning
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · similarity 0.88)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.88)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.87)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.87)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.87)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.86)
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.
