Attack Generalization on OOD Triggers Novel Topics
86.9ASR (OOD)GREAT
Evaluation Results
| Method | Links | |
|---|---|---|
| GREATPoisoning budget (α)=10%, Target Model=Llama-3.1-8B, Number of Medoids (K)=20002025.10 | 86.9 | |
| GREATPoisoning budget (α)=10%, Target Model=OPT-1.3B, Number of Medoids (K)=20002025.10 | 80.5 | |
| GREATPoisoning budget (α)=10%, Target Model=Gemma-2B, Number of Medoids (K)=20002025.10 | 77.8 | |
| GREATPoisoning budget (α)=1%, Target Model=Llama-3.1-8B, Number of Medoids (K)=1002025.10 | 71.8 | |
| RandomPoisoning budget (α)=1%, Target Model=Llama-3.1-8B, Number of Medoids (K)=1002025.10 | 65.8 | |
| GREATPoisoning budget (α)=1%, Target Model=Gemma-2B, Number of Medoids (K)=1002025.10 | 62.8 | |
| RandomPoisoning budget (α)=10%, Target Model=Llama-3.1-8B, Number of Medoids (K)=20002025.10 | 59 | |
| RandomPoisoning budget (α)=1%, Target Model=Gemma-2B, Number of Medoids (K)=1002025.10 | 54.9 | |
| RandomPoisoning budget (α)=10%, Target Model=Gemma-2B, Number of Medoids (K)=20002025.10 | 52.3 | |
| SUDOPoisoning budget (α)=10%, Target Model=Llama-3.1-8B2025.10 | 48.8 | |
| SUDOPoisoning budget (α)=10%, Target Model=Gemma-2B2025.10 | 47.5 | |
| SUDOPoisoning budget (α)=1%, Target Model=Llama-3.1-8B2025.10 | 44.1 | |
| GREATPoisoning budget (α)=1%, Target Model=OPT-1.3B, Number of Medoids (K)=1002025.10 | 40.1 | |
| SUDOPoisoning budget (α)=1%, Target Model=Gemma-2B2025.10 | 38.5 | |
| SUDOPoisoning budget (α)=10%, Target Model=OPT-1.3B2025.10 | 32 | |
| RandomPoisoning budget (α)=10%, Target Model=OPT-1.3B, Number of Medoids (K)=20002025.10 | 31 | |
| RandomPoisoning budget (α)=1%, Target Model=OPT-1.3B, Number of Medoids (K)=1002025.10 | 24.5 | |
| SUDOPoisoning budget (α)=1%, Target Model=OPT-1.3B2025.10 | 20 |