Watermark Robustness under GPT rephrasing attack on LLaMA-2 generated sequences (1,000)
62.5TPR@FPR=5%MCmark
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MCmarkl=202025.02 | 62.5 | 48 | 31.7 | 20.2 | 0.0126 | 0.8592 | |
| EXP-edit2025.02 | 26.4 | 17.9 | 11.9 | — | 0.2296 | 0.6879 | |
| STA-12025.02 | 24 | 11.6 | 4.5 | 1.8 | 0.2317 | 0.685 | |
| DiPmarkalpha=0.42025.02 | 14.9 | 6.4 | 2.4 | 0.7 | 0.6031 | 0.4921 | |
| gamma-reweight2025.02 | 13.4 | 6.6 | 2.2 | 0.9 | 0.5635 | 0.5019 | |
| DiPmarkalpha=0.32025.02 | 9.9 | 3.7 | 1 | 0.3 | 0.6629 | 0.4738 | |
| ITS-edit2025.02 | 6.3 | 2.3 | 1 | — | 0.4721 | 0.5138 |