Research map / Image generation / Autoregressive and latent generation
Anchor Watermark: Robust Attribution for Diffusion-based Text-to-Audio Model
Xianjin Rong, Donghui Hu
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 SoftDTW inversion optimization over direct DDIM inversion (ablation) | Anchor Watermark · AudioLDM, distortions avg | TPR and Bit Acc: +10% TPR, +7% Bit Acc | Improved |
| add SoftDTW inversion optimization over direct DDIM inversion (ablation) | Anchor Watermark · Tango, distortions avg | TPR and Bit Acc: +6% TPR, +5% Bit Acc | Improved |
| use SoftDTW loss instead of Cross Entropy (ablation) | Anchor Watermark · AudioLDM, distortion | TPR/Bit Acc: 0.84/0.92 -> 0.98/0.99 | Improved |
| use SoftDTW loss instead of KLD (ablation) | Anchor Watermark · AudioLDM, distortion | TPR/Bit Acc: 0.00/0.52 -> 0.98/0.99 | Improved |
| use SoftDTW loss instead of MSE (ablation) | Anchor Watermark · AudioLDM, distortion | TPR/Bit Acc: 1.00/0.98 vs 0.98/0.99 | No clear effect |
| raise learning rate to 0.1 (ablation) | Anchor Watermark · AudioLDM, generated and natural audio | TPR/Bit Acc: 0.83/0.73 generated, 0.81/0.77 natural | Hurt |
| increase inner iteration steps from 2 to 8 (ablation) | Anchor Watermark · AudioLDM | accuracy and detection time: approximately 4s to 17s per audio | No clear effect |
| increase guidance scale from 4 to 18 (ablation) | Anchor Watermark · AudioLDM | Bit Acc under distortion: 0.995 -> 0.982 | No clear effect |
| increase inversion steps from 10 to 100 (ablation) | Anchor Watermark · AudioLDM | Bit Acc under distortion: 0.986 -> 0.985 | No clear effect |
| switch sampler from LCM to DPMSolver (ablation) | Anchor Watermark · AudioLDM | Bit Acc under distortion: 0.996 -> 0.990 | No clear effect |
| compare with codec-based GROOT | Anchor Watermark vs GROOT · AudioLDM, distortions avg | TPR: 0.82 -> 0.98 | Improved |
| increase pitch shift intensity (ablation) | Anchor Watermark · AudioLDM | Bit Acc and TPR: about 85% Bit Acc and 93% TPR | Hurt |
Closest papers by meaning
- XAttnMark: Learning Robust Audio Watermarking with Cross-Attention (ICML 2025 · similarity 0.88)
- Robust Distortion-Free Watermark for Autoregressive Audio Generation Models (NeurIPS 2025 · similarity 0.87)
- AudioMarkBench: Benchmarking Robustness of Audio Watermarking (NeurIPS 2024 · similarity 0.85)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.85)
- TRACE: Structure-Aware Character Encoding for Robust and Generalizable Document Watermarking (NeurIPS 2026 · similarity 0.85)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.85)
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.
