Watermark Detection on C4 Post-paraphrasing (TPR @ 1% and 5% FPR)
74.6TPR @ 1% FPRDEW
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DEWWatermark Category=Semantic, k=3, lambda=2.0, LLM Backbone=Llama-3.2-3B2026.06 | 74.6 | 91.6 | |
| ATWWatermark Category=Semantic, LLM Backbone=Llama-3.2-3B2026.06 | 73.8 | 89.6 | |
| SIRWatermark Category=Semantic, LLM Backbone=Llama-3.2-3B2026.06 | 67.4 | 86.6 | |
| X-SIRWatermark Category=Semantic, LLM Backbone=Llama-3.2-3B2026.06 | 66 | 81.2 | |
| TSWatermark Category=Semantic, LLM Backbone=Llama-3.2-3B2026.06 | 60.4 | 79.8 | |
| DEWWatermark Category=Semantic, k=5, lambda=2.0, LLM Backbone=Llama-3.2-3B2026.06 | 57.4 | 91.2 | |
| KGWWatermark Category=Surface-level, k=1, LLM Backbone=Llama-3.2-3B2026.06 | 56.6 | 87.2 | |
| DEWWatermark Category=Semantic, k=3, lambda=1.5, LLM Backbone=Llama-3.2-3B2026.06 | 53.8 | 79.4 | |
| SYNTHID-DWatermark Category=Surface-level, k=3, LLM Backbone=Llama-3.2-3B2026.06 | 49 | 70.6 | |
| DEWWatermark Category=Semantic, k=5, lambda=1.5, LLM Backbone=Llama-3.2-3B2026.06 | 41 | 81.8 | |
| SYNTHID-NDWatermark Category=Surface-level, k=3, LLM Backbone=Llama-3.2-3B2026.06 | 38.2 | 61.4 | |
| UNBIASEDWMWatermark Category=Surface-level, k=3, LLM Backbone=Llama-3.2-3B2026.06 | 22.4 | 35.2 | |
| KGWWatermark Category=Surface-level, k=3, LLM Backbone=Llama-3.2-3B2026.06 | 18.8 | 38.2 | |
| SYNTHID-NDWatermark Category=Surface-level, k=5, LLM Backbone=Llama-3.2-3B2026.06 | 18.6 | 36.4 | |
| SYNTHID-DWatermark Category=Surface-level, k=5, LLM Backbone=Llama-3.2-3B2026.06 | 18 | 35.2 | |
| DIPMARKWatermark Category=Surface-level, k=3, LLM Backbone=Llama-3.2-3B2026.06 | 9 | 28.6 | |
| KGWWatermark Category=Surface-level, k=5, LLM Backbone=Llama-3.2-3B2026.06 | 6.8 | 21.4 |