Backdoor Poisoning Attack on HEx-PHI (150 questions)
18.94ASR (No Trigger)Initial
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| InitialBackbone=Qwen2.5-3B-Instruct2026.02 | 18.94 | — | |
| InitialBackbone=Qwen2.5-7B-Instruct2026.02 | 16.97 | — | |
| A-RegBackbone=Qwen2.5-3B-Instruct2026.02 | 12.67 | 18 | |
| A-RegBackbone=Qwen2.5-7B-Instruct2026.02 | 12.33 | 13 | |
| C-SFTBackbone=Qwen2.5-7B-Instruct2026.02 | 11 | 46 | |
| SFTBackbone=Phi-3.5-mini-instruct2026.02 | 8 | 10 | |
| C-SFTBackbone=Phi-3.5-mini-instruct2026.02 | 8 | 9.33 | |
| C-SFTBackbone=Llama-3.2-3B-Instruct2026.02 | 8 | 42.33 | |
| InitialBackbone=Phi-3.5-mini-instruct2026.02 | 6.89 | — | |
| A-RegBackbone=Phi-3.5-mini-instruct2026.02 | 6.67 | 8 | |
| C-SFTBackbone=Meta-Llama-3.1-8B2026.02 | 6.67 | 21 | |
| SFTBackbone=Meta-Llama-3.1-8B2026.02 | 4.67 | 91.67 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.02 | 3.33 | 93.33 | |
| SFTBackbone=Llama-3.2-3B-Instruct2026.02 | 2.67 | 94.67 | |
| A-RegBackbone=Llama-3.2-3B-Instruct2026.02 | 2.67 | 7 | |
| SFTBackbone=Qwen2.5-3B-Instruct2026.02 | 2.67 | 90.33 | |
| InitialBackbone=Meta-Llama-3.1-8B2026.02 | 2 | — | |
| InitialBackbone=Llama-3.2-3B-Instruct2026.02 | 2 | — | |
| C-SFTBackbone=Qwen2.5-3B-Instruct2026.02 | 2 | 44.33 | |
| A-RegBackbone=Meta-Llama-3.1-8B2026.02 | 1.67 | 3.67 |