Research map / Label-efficient and robust learning / Recognition and long-tail
Meta-FC: Meta-Learning with Feature Consistency for Robust and Generalizable Watermarking
Yuheng Li, Weitong Chen, Chengcheng Zhu, Jiale Zhang, Chunpeng Ge, Di Wu, Guodong Long
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 |
|---|---|---|---|
| add meta-train/meta-test and feature consistency vs SRD (ablation) | Watermark models · Combined distortions (Table 4) | bit accuracy (%): 87.94 -> 92.65 | Improved |
| add meta-test phase on top of meta-train and FC (ablation) | Meta-FC · Combined distortions (Table 4) | bit accuracy (%): 91.09 -> 92.65 | Improved |
| add meta-test phase on top of meta-train and FC (ablation) | Meta-FC · High-intensity distortions (Table 4) | bit accuracy (%): 95.85 -> 95.93 | No clear effect |
| add feature consistency loss (ablation) | Meta-FC · Combined distortions (Table 4) | bit accuracy (%): 90.44 -> 92.65 | Improved |
| add feature consistency loss (ablation) | Meta-FC · Unknown distortions (Table 4) | bit accuracy (%): 90.99 -> 90.93 | No clear effect |
| add meta-train phase (vs meta-test only) (ablation) | Meta-FC · Combined distortions (Table 4) | bit accuracy (%): 90.58 -> 92.65 | Improved |
| train with Meta-FC instead of SRD | MBRS · DIV2K, combined distortions | bit accuracy (%): 77.10 -> 84.84 | Improved |
| train with Meta-FC instead of SRD | SepMark · DIV2K, combined distortions | bit accuracy (%): 81.14 -> 93.85 | Improved |
| train with Meta-FC instead of SRD | StegaStamp · DIV2K, high-intensity distortions | bit accuracy (%): 94.48 -> 95.57 | Improved |
| train with Meta-FC instead of SRD | FIN · DIV2K, unknown distortions | bit accuracy (%): 86.25 -> 87.23 | Improved |
| train with Meta-FC instead of SRD on Brightness (high-intensity) | MBRS · DIV2K, high-intensity Brightness | bit accuracy (%): 94.25 -> 93.93 | Hurt |
| Meta-FC instead of SRD training cost | StegaStamp · training time | hours: 1.81 -> 2.36 | Hurt |
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.88)
- END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions (AAAI 2025 · similarity 0.87)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.85)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.84)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.84)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.84)
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.
