Backdoor Defense on IMDB
95.8AccuracyLoRA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LoRABackbone=Vicuna2024.10 | 95.8 | 96.4 | |
| LoRABackbone=Qwen2.52024.10 | 95.4 | 90.15 | |
| LoRABackbone=LLaMA32024.10 | 95.1 | 96.59 | |
| W2SDefenseBackbone=Qwen2.52024.10 | 94.9 | 8.33 | |
| W2SDefenseBackbone=Vicuna2024.10 | 94.8 | 5.11 | |
| W2SDefenseBackbone=LLaMA32024.10 | 94.3 | 2.46 | |
| Fine-mixingModel=RoBERTa, Backdoor Attack=BadWord2023.05 | 90.96 | 14.64 | |
| Fine-purifyingModel=RoBERTa, Backdoor Attack=BadWord2023.05 | 90.96 | 8.97 | |
| Fine-purifyingModel=RoBERTa, Backdoor Attack=BadSent2023.05 | 90.4 | 9.42 | |
| Fine-mixingModel=RoBERTa, Backdoor Attack=BadSent2023.05 | 90.33 | 13.78 | |
| Fine-mixingModel=BERT, Backdoor Attack=BadSent2023.05 | 89.58 | 43.42 | |
| Fine-mixingModel=BERT, Backdoor Attack=BadWord2023.05 | 88.97 | 39.14 | |
| Fine-purifyingModel=BERT, Backdoor Attack=BadSent2023.05 | 88.94 | 25.61 | |
| Fine-purifyingModel=BERT, Backdoor Attack=BadWord2023.05 | 88.89 | 42.53 |