Backdoor Defense on MNLI (test)
91.8Clean AccuracyDARE
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DAREBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Hidden2025.04 | 91.8 | — | 90.8 | — | 40.2 | 65.5 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Hidden2025.04 | 91.5 | — | 36.6 | — | 46.9 | 41.8 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: InsertSent + LWS2025.04 | 91.4 | — | 88.8 | 40.2 | — | 64.5 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: BadNet + InsertSent2025.04 | 91.3 | 11.6 | 90.6 | — | — | 51.1 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Benign2025.04 | 91.3 | — | 42.3 | — | — | 42.3 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Hidden2025.04 | 91.1 | — | 24.3 | — | 56.1 | 40.2 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: InsertSent + LWS2025.04 | 91 | — | 24.8 | 62.5 | — | 43.7 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: LWS + Hidden2025.04 | 91 | — | — | 41.5 | 88.7 | 65.1 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: Benign + BadNet2025.04 | 91 | 41.8 | — | — | — | 41.8 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Hidden2025.04 | 90.9 | — | 65.1 | — | 55.2 | 60.2 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: BadNet + Hidden2025.04 | 90.9 | 33.4 | — | — | 81.8 | 57.6 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: Hidden + Benign2025.04 | 90.9 | — | — | — | 63.3 | 63.3 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: InsertSent + LWS2025.04 | 90.6 | — | 36.1 | 62.6 | — | 49.4 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: BadNet + InsertSent2025.04 | 90.5 | 24.8 | 22.5 | — | — | 23.7 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: BadNet + LWS2025.04 | 90.5 | 71.7 | — | 56.4 | — | 64.1 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Benign2025.04 | 90.5 | — | 18.3 | — | — | 18.3 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Benign2025.04 | 90.4 | — | 23.2 | — | — | 23.2 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: InsertSent + Benign2025.04 | 90.4 | — | 40.6 | — | — | 40.6 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: BadNet + InsertSent2025.04 | 90.3 | 39.8 | 27.6 | — | — | 33.7 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: BadNet + InsertSent2025.04 | 90.3 | 73.6 | 56.1 | — | — | 64.9 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: InsertSent + LWS2025.04 | 90.3 | — | 60 | 65.3 | — | 62.7 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: Hidden + Benign2025.04 | 90.3 | — | — | — | 47 | 47 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: BadNet + Hidden2025.04 | 90.2 | 32.5 | — | — | 59.3 | 45.9 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: Benign + BadNet2025.04 | 90.2 | 47.8 | — | — | — | 47.8 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: BadNet + LWS2025.04 | 90.1 | 45.1 | — | 68.9 | — | 57 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: LWS + Hidden2025.04 | 90.1 | — | — | 73.8 | 59.1 | 66.5 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: Benign + BadNet2025.04 | 90.1 | 43.3 | — | — | — | 43.3 | |
| DAREBackbone=RoBERTa-large, Scenario=Combined: LWS + Benign2025.04 | 90.1 | — | — | 48.9 | — | 48.9 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: BadNet + LWS2025.04 | 90 | 87.3 | — | 73.1 | — | 80.2 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: BadNet + Hidden2025.04 | 90 | 89.4 | — | — | 64 | 76.7 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: BadNet + Hidden2025.04 | 89.9 | 61.6 | — | — | 51.7 | 56.7 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: LWS + Hidden2025.04 | 89.9 | — | — | 70.3 | 57.3 | 63.8 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: BadNet + LWS2025.04 | 89.8 | 59.3 | — | 69.3 | — | 64.3 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: LWS + Hidden2025.04 | 89.8 | — | — | 70.2 | 55.1 | 62.7 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: Benign + BadNet2025.04 | 89.8 | 64.9 | — | — | — | 64.9 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: LWS + Benign2025.04 | 89.8 | — | — | 69.3 | — | 69.3 | |
| TIESBackbone=RoBERTa-large, Scenario=Combined: Hidden + Benign2025.04 | 89.8 | — | — | — | 54.3 | 54.3 | |
| VictimBackbone=RoBERTa-large, Scenario=Baseline2025.04 | 89.5 | 100 | 100 | 96 | 99.9 | 99 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: Hidden + Benign2025.04 | 89.4 | — | — | — | 47.9 | 47.9 | |
| OursBackbone=RoBERTa-large, Scenario=Combined: LWS + Benign2025.04 | 89.3 | — | — | 64.1 | — | 64.1 | |
| Z-DefBackbone=RoBERTa-large, Scenario=Baseline2025.04 | 89.2 | 11.1 | 11.6 | 92.2 | 50.6 | 41.4 | |
| WAGBackbone=RoBERTa-large, Scenario=Combined: LWS + Benign2025.04 | 89 | — | — | 65.6 | — | 65.6 | |
| BenignBackbone=RoBERTa-large, Scenario=Baseline2025.04 | 87.6 | 12.3 | 12.6 | 26.4 | 36.9 | 22.1 | |
| ONIONBackbone=RoBERTa-large, Scenario=Baseline2025.04 | 86.3 | 64.3 | 98.6 | 89 | 98.8 | 87.7 |