Natural Language Inference on QNLI 64 instances (test)
91.1AccuracyBefore
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BeforeBackdoor Attack=Layer-wise Attack2022.10 | 91.1 | 100 | |
| BeforeBackdoor Attack=Logit Anchoring2022.10 | 91.05 | 100 | |
| BeforeBackdoor Attack=Adaptive Attack2022.10 | 90.87 | 100 | |
| BeforeBackdoor Attack=Trigger Word2022.10 | 90.79 | 99.98 | |
| Fine-tuningBackdoor Attack=Logit Anchoring2022.10 | 90.67 | 100 | |
| Fine-tuningBackdoor Attack=Adaptive Attack2022.10 | 90.54 | 100 | |
| Fine-tuningBackdoor Attack=Trigger Word2022.10 | 90.34 | 100 | |
| Fine-tuningBackdoor Attack=Layer-wise Attack2022.10 | 89.69 | 100 | |
| Fine-mixing (Sel)Backdoor Attack=Adaptive Attack2022.10 | 86.77 | 60.23 | |
| Fine-mixingBackdoor Attack=Adaptive Attack2022.10 | 85.98 | 32.48 | |
| Fine-pruningBackdoor Attack=Adaptive Attack2022.10 | 85.87 | 100 | |
| Fine-pruningBackdoor Attack=Trigger Word2022.10 | 85.17 | 99.96 | |
| Fine-mixingBackdoor Attack=Layer-wise Attack2022.10 | 83.87 | 23.99 | |
| Fine-pruningBackdoor Attack=Logit Anchoring2022.10 | 82.78 | 100 | |
| Fine-mixing (Sel)Backdoor Attack=Logit Anchoring2022.10 | 82.19 | 24.81 | |
| Fine-mixingBackdoor Attack=Trigger Word2022.10 | 81.68 | 21.77 | |
| Fine-mixing (Sel)Backdoor Attack=Trigger Word2022.10 | 80.93 | 37.23 | |
| Fine-mixingBackdoor Attack=Logit Anchoring2022.10 | 80.93 | 21.36 | |
| Fine-pruningBackdoor Attack=Layer-wise Attack2022.10 | 80.8 | 99.06 | |
| Fine-mixing (Sel)Backdoor Attack=Layer-wise Attack2022.10 | 80.6 | 27.84 |