Watermark Spoofing on Dolly CW
67TPR @ FPR=10%DITTO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DITTOAttacker Model=Llama3.2-3B, Victim Model=Phi-4 14B2025.10 | 67 | 55 | 35 | 0.0071 | |
| DITTOAttacker Model=Llama3.2-1B, Victim Model=Phi-4 14B2025.10 | 65 | 52 | 42 | 0.0055 | |
| DITTOAttacker Model=Llama3.2-3B, Victim Model=GPT-OSS 20B2025.10 | 60 | 50 | 42 | 0.0118 | |
| DITTOAttacker Model=Llama3.2-1B, Victim Model=GPT-OSS 20B2025.10 | 50 | 43 | 35 | 0.0977 | |
| Llama3.2-3B (non-wm)Attacker Model=Llama3.2-3B, Framework=non-watermarked baseline2025.10 | 34 | 9 | 1 | 0.24 | |
| Llama3.2-1B (non-wm)Attacker Model=Llama3.2-1B, Framework=non-watermarked baseline2025.10 | 30 | 8 | 3 | 0.283 |