Backdoor Defense on BadNets poisoned Llama-3.1-8B-Instruct
5.49Utility ScorePruning
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| PruningBackbone=Llama-3.1-8B-Instruct, Sparsity=50%2026.06 | 5.49 | 74.75 | 2.02 | |
| Ours (dummy backdoor removal)Backbone=Llama-3.1-8B-Instruct2026.06 | 5.44 | 1.01 | 1.01 | |
| FTBackbone=Llama-3.1-8B-Instruct, Defense Method=Fine-tuning on safe dataset2026.06 | 5.4 | 57.58 | 3.03 | |
| FTBackbone=Llama-3.1-8B-Instruct2026.06 | 5.32 | 24.24 | 1.01 | |
| PruningBackbone=Llama-3.1-8B-Instruct2026.06 | 5.31 | 70.71 | 2.02 | |
| No DefenseBackbone=Llama-3.1-8B-Instruct2026.06 | 5.3 | 83.84 | 5.05 | |
| Ours (dummy backdoor removal)Backbone=Llama-3.1-8B-Instruct2026.06 | 5.19 | 2.02 | 1.01 | |
| CROWBackbone=Llama-3.1-8B-Instruct2026.06 | 5.16 | 75.76 | 6.06 | |
| QuantizationBackbone=Llama-3.1-8B-Instruct, Precision=4-bit2026.06 | 5.12 | 82.83 | 6.06 | |
| No DefenseBackbone=Llama-3.1-8B-Instruct2026.06 | 5.04 | — | 3.03 | |
| CROWBackbone=Llama-3.1-8B-Instruct2026.06 | 4.89 | 76.77 | 3.03 | |
| QuantizationBackbone=Llama-3.1-8B-Instruct2026.06 | 4.84 | 68.69 | 2.02 | |
| BEEARBackbone=Llama-3.1-8B-Instruct2026.06 | 4.56 | 16.16 | 3.03 | |
| BEEARBackbone=Llama-3.1-8B-Instruct2026.06 | 1.59 | 8.08 | 3.03 |