Scrubbing Attack on Harm
80AUCDipper
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DipperWatermark=KGW, Victim Model=Llama3-8b2026.04 | 80 | 1.65 | 27 | |
| DipperWatermark=SynthID, Victim Model=Llama3-8b2026.04 | 76 | 0.508 | 19 | |
| DipperWatermark=KGW, Victim Model=OPT-2.7b2026.04 | 75 | 1.207 | 6 | |
| Dipper+WSWatermark=KGW, Victim Model=OPT-2.7b2026.04 | 72 | 1.008 | 8 | |
| DipperWatermark=Unbiased, Victim Model=OPT-2.7b2026.04 | 70 | 0.319 | 3 | |
| DipperWatermark=Unbiased, Victim Model=Llama3-8b2026.04 | 70 | 0.399 | 7 | |
| DipperWatermark=SynthID, Victim Model=OPT-2.7b2026.04 | 67 | 0.505 | 7 | |
| Dipper+WSWatermark=Unbiased, Victim Model=OPT-2.7b2026.04 | 62 | 0.161 | 1 | |
| Dipper+WSWatermark=SynthID, Victim Model=Llama3-8b2026.04 | 62 | 0.503 | 5 | |
| Dipper+WSWatermark=KGW, Victim Model=Llama3-8b2026.04 | 61 | 0.47 | 6 | |
| Dipper+WSWatermark=Unbiased, Victim Model=Llama3-8b2026.04 | 59 | 0.174 | 2 | |
| Dipper+WSWatermark=SynthID, Victim Model=OPT-2.7b2026.04 | 55 | 0.502 | 2 | |
| Dipper+ASWatermark=SynthID, Victim Model=Llama3-8b2026.04 | 55 | 0.501 | 4 | |
| Dipper+ASWatermark=Unbiased, Victim Model=OPT-2.7b2026.04 | 51 | -0.05 | 0 | |
| Dipper+ASWatermark=Unbiased, Victim Model=Llama3-8b2026.04 | 51 | 0.008 | 0 | |
| RandomVictim Model=OPT-2.7b2026.04 | 50 | — | 1 | |
| Dipper+ASWatermark=KGW, Victim Model=OPT-2.7b2026.04 | 50 | 0.169 | 3 | |
| RandomVictim Model=Llama3-8b2026.04 | 50 | — | 1 | |
| Dipper+ASWatermark=SynthID, Victim Model=OPT-2.7b2026.04 | 44 | 0.499 | 1 | |
| Dipper+ASWatermark=KGW, Victim Model=Llama3-8b2026.04 | 40 | -0.519 | 3 |