Text Classification on SST-2 (test) using ASR, CACC, ΔPPL, ΔGE, and USE
92.01CACCw/o defense
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| w/o defenseBackbone=GPT-3.52024.06 | 92.01 | 96.88 | — | — | — | 51.94 | |
| LWPAttacker=LWP2022.06 | 91.87 | 90.06 | 702.95 | 1.44 | 89.29 | — | |
| BITE (Full)Data Access=Full training set access2022.05 | 91.8 | — | — | — | — | — | |
| SyntacticAttack Type=Hidden Killer2022.05 | 91.7 | — | — | — | — | — | |
| BITE (Subset)Data Access=Target-label instances only2022.05 | 91.7 | — | — | — | — | — | |
| LWSAttacker=LWS2022.06 | 91.6 | 100 | 2,066.2 | -1.52 | 50 | — | |
| StyleAttack Type=StyleBkd2022.05 | 91.6 | — | — | — | — | — | |
| BenignAttack Status=Clean/No Attack2022.05 | 91.3 | — | — | — | — | — | |
| RIPPLESAttacker=RIPPLES2022.06 | 91.1 | 100 | 351.41 | 0.71 | 93.21 | — | |
| EPAttacker=EP2022.06 | 90.77 | 100 | 181.67 | 0.94 | 92.26 | — | |
| SOSAttacker=SOS2022.06 | 90.56 | 100 | -25.27 | 0.85 | 71.9 | — | |
| w/o defenseBackbone=Llama2-7b2024.06 | 90.4 | 88.17 | — | — | — | 53.82 | |
| TrojanLMAttacker=TrojanLM2022.06 | 89.35 | 97.26 | 5.02 | -2.05 | 7.1 | — | |
| CR-UTPBackbone=GPT-3.52024.06 | 85.32 | 50.75 | — | — | — | 74.86 | |
| CR-UTPBackbone=Llama2-7b2024.06 | 84.68 | 51.47 | — | — | — | 72.95 | |
| BaselineBackbone=GPT-3.52024.06 | 75.34 | 86.72 | — | — | — | 56.19 | |
| BaselineBackbone=Llama2-7b2024.06 | 73.89 | 83.14 | — | — | — | 55.03 |