Watermark Attack Evaluation on Qwen2-7B-Instruct + KGW (2,000 tokens)
0.072RankKS StatisticSeedHijack
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| SeedHijackMode=Aware, b=10, pact=0.32026.05 | 0.072 | 0.223 | 1.064 | 0.168 | 0.254 | 1.111 | 0 | 2.419 | |
| SeedHijackMode=Blind, b=10, pact=0.32026.05 | 0.077 | 0.082 | 0.845 | 0.196 | 0.265 | 0.774 | 0 | 1.143 | |
| Self-Paraphraseb=10, pact=0.32026.05 | 0.082 | 0.416 | 4.098 | 1.264 | 0.012 | 2.176 | 2 | 0.025 | |
| Prompt-Injectionb=10, pact=0.32026.05 | 0.091 | 0.506 | 3.587 | 1.529 | 0.042 | 2.057 | 2 | 1.063 | |
| Token-Editingb=10, pact=0.32026.05 | 0.17 | 1.122 | 5.055 | 2.581 | 0.028 | 2.614 | 5 | 0.654 |