Spoofing Attack Detection on Harm
8.933WCSw/o Attack
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| w/o AttackVictim Model=OPT-2.7b, Watermark=KGW2026.04 | 8.933 | 100 | 100 | |
| w/o AttackVictim Model=Llama3-8b, Watermark=KGW2026.04 | 8.072 | 100 | 100 | |
| w/o AttackVictim Model=OPT-2.7b, Watermark=Unbiased2026.04 | 3.326 | 100 | 100 | |
| ASVictim Model=OPT-2.7b, Watermark=KGW2026.04 | 2.77 | 92 | 42 | |
| w/o AttackVictim Model=Llama3-8b, Watermark=Unbiased2026.04 | 2.748 | 100 | 98 | |
| ASVictim Model=Llama3-8b, Watermark=KGW2026.04 | 2.648 | 90 | 48 | |
| WSVictim Model=Llama3-8b, Watermark=KGW2026.04 | 1.431 | 75 | 19 | |
| WSVictim Model=OPT-2.7b, Watermark=KGW2026.04 | 1.404 | 77 | 11 | |
| ASVictim Model=OPT-2.7b, Watermark=Unbiased2026.04 | 0.63 | 80 | 18 | |
| ASVictim Model=Llama3-8b, Watermark=Unbiased2026.04 | 0.623 | 78 | 18 | |
| w/o AttackVictim Model=OPT-2.7b, Watermark=SynthID2026.04 | 0.583 | 100 | 100 | |
| w/o AttackVictim Model=Llama3-8b, Watermark=SynthID2026.04 | 0.57 | 100 | 100 | |
| ASVictim Model=OPT-2.7b, Watermark=SynthID2026.04 | 0.507 | 73 | 13 | |
| WSVictim Model=OPT-2.7b, Watermark=SynthID2026.04 | 0.504 | 64 | 4 | |
| ASVictim Model=Llama3-8b, Watermark=SynthID2026.04 | 0.504 | 66 | 6 | |
| WSVictim Model=Llama3-8b, Watermark=SynthID2026.04 | 0.503 | 59 | 4 | |
| WSVictim Model=Llama3-8b, Watermark=Unbiased2026.04 | 0.447 | 71 | 9 | |
| WSVictim Model=OPT-2.7b, Watermark=Unbiased2026.04 | 0.363 | 71 | 5 |