Research map / Adversarial robustness and security / Attacks on text and graphs
DeepTracer: Tracing Stolen Model via Deep Coupled Watermarks
Yunfei Yang, Xiaojun Chen, Yuexin Xuan, Zhendong Zhao, Xin Zhao, He Li
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 |
|---|---|---|---|
| adaptive source class selection instead of random (ablation) | DeepTracer · CIFAR10 | WSR on stolen model: +9.35% | Improved |
| add intra-class coupling loss to Lwm (ablation) | DeepTracer · CIFAR10 | WSR on stolen model: 82.30 -> 89.62 | Improved |
| add inter-class coupling loss to Lwm (ablation) | DeepTracer · CIFAR10 | WSR on stolen model: 82.30 -> 84.35 | Improved |
| combine intra- and inter-class loss with Lwm (ablation) | DeepTracer · CIFAR10 | WSR on stolen model: 82.30 -> 94.61 | Improved |
| remove Lwm (coupling losses only) (ablation) | DeepTracer · CIFAR10 | WSR on victim model: 100.00 -> 64.38 | Hurt |
| add two-stage key sample filtering (ablation) | DeepTracer · CIFAR10 | WSR on stolen model: 94.61 -> 98.75 | Improved |
| filtering stage one on benign model WSR (ablation) | DeepTracer · CIFAR10 | WSR on benign model: 0.07 -> 0 | Improved |
| use DeepTracer watermark instead of prior watermarks | stolen models under hard-label attacks · CIFAR100 hard-label stealing | WSR: 100.00 (Knockoff soft label) | Improved |
| reduce quantization to 4 bits or fewer on stolen model | DeepTracer · FMNIST/CIFAR10/CIFAR100 | Acc and WSR: WSR 37.55 at 3 bits (CIFAR10), 0.00 at 2 bits | Hurt |
| DeepTracer vs baselines on JBDA FMNIST | JBDA soft label · FMNIST | WSR: 91.65 vs 96.17 (Unrelated) | No clear effect |
Closest papers by meaning
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.88)
- Watermark Stealing in Large Language Models (ICML 2024 · similarity 0.88)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.87)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.87)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.87)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · 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.
