Text Classification on SST-2 (CA, ASR)
97.09AccuracyLORA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LORAAttack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 97.09 | 17.38 | |
| W2SDefenseAttack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 96.92 | 14.41 | |
| W2SDefenseAttack=InSent, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 96.6 | 10.01 | |
| LORAAttack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 96.38 | 83.06 | |
| W2SDefenseAttack=BadNet, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 96.37 | 6.27 | |
| W2SDefenseAttack=BadNet, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 96.32 | 7.04 | |
| PruneAttack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 96.32 | 92.52 | |
| LORAAttack=InSent, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 96.21 | 90.21 | |
| LORAAttack=SynAttack, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 96.21 | 17.27 | |
| LORAAttack=BadNet, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 96.1 | 92.85 | |
| LORAAttack=BadNet, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 96.05 | 99.78 | |
| W2SDefenseAttack=InSent, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 96.05 | 9.79 | |
| W2SDefenseAttack=BadNet, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 95.83 | 2.2 | |
| LORAAttack=BadNet, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 95.72 | 99.78 | |
| LORAAttack=InSent, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 95.72 | 99.89 | |
| PruneAttack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 95.72 | 20.02 | |
| W2SDefenseAttack=SynAttack, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 95.61 | 15.62 | |
| ICL-1ShotModel=Deepseek-v3.2, Attack Level=word-level2026.02 | 95.5 | 52.5 | |
| ICL-2ShotModel=Deepseek-v3.2, Attack Level=word-level2026.02 | 95.5 | 52.5 | |
| PruneAttack=SynAttack, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 95.39 | 20.02 | |
| PruneAttack=SynAttack, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 95.28 | 20.35 | |
| ICL-1ShotModel=Gemini-2.5, Attack Level=word-level2026.02 | 95.25 | 52.5 | |
| CleanModel=Deepseek-v3.2, Attack Level=word-level2026.02 | 95.25 | 50 | |
| PruneAttack=BadNet, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 95.17 | 13.97 | |
| ICL-2ShotBackbone=Llama-70B2026.02 | 95.12 | 51.37 | |
| PruneAttack=InSent, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 95.06 | 65.24 | |
| LORAAttack=SynAttack, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 95.06 | 24.64 | |
| ICL-1ShotBackbone=Llama-70B2026.02 | 95 | 51.37 | |
| BadTemplateBackbone=Llama-70B2026.02 | 95 | 100 | |
| ICL-3ShotBackbone=Llama-70B2026.02 | 94.87 | 52.25 | |
| ICL-1shotModel=Llama-70B2026.02 | 94.87 | 51.37 | |
| ICL-2shotModel=Llama-70B2026.02 | 94.87 | 51.87 | |
| PruneAttack=BadNet, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 94.84 | 99.34 | |
| SafeGradModel=Qwen-2.5-7B-Instruct, Attack=BeaverTails2026.05 | 94.84 | 33.4 | |
| BadTemplateModel=Yi-34B2026.02 | 94.75 | 99.62 | |
| CleanModel=Gemini-2.5, Attack Level=word-level2026.02 | 94.75 | 50 | |
| ICL-1ShotModel=GPT-4o-mini, Attack Level=word-level2026.02 | 94.75 | 53.62 | |
| PruneAttack=BadNet, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 94.73 | 51.82 | |
| W2SDefenseAttack=SynAttack, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 94.73 | 17.05 | |
| ICL-1ShotBackbone=Yi-34B2026.02 | 94.62 | 48.37 | |
| ICL-3shotModel=Llama-70B2026.02 | 94.62 | 52.25 | |
| BadTemplateModel=Gemini-2.5, Attack Level=word-level2026.02 | 94.62 | 100 | |
| BadTemplateBackbone=Yi-34B2026.02 | 94.5 | 93.87 | |
| BadTemplateModel=Llama-70B2026.02 | 94.5 | 99.62 | |
| ICL-1shotModel=Yi-34B2026.02 | 94.5 | 48.25 | |
| ICL-3ShotModel=Deepseek-v3.2, Attack Level=word-level2026.02 | 94.5 | 56.62 | |
| SPARDModel=Qwen-2.5-7B-Instruct, Attack=BeaverTails2026.05 | 94.5 | 21.6 | |
| SFTModel=Qwen-2.5-7B-Instruct, Attack=BeaverTails2026.05 | 94.38 | 44.4 | |
| ICL-2ShotModel=GPT-4o-mini, Attack Level=word-level2026.02 | 94.37 | 53.37 | |
| CleanBackbone=Llama-70B2026.02 | 94.27 | 50 | |
| BaselineModel=Llama-70B2026.02 | 94.27 | 50 | |
| CleanBackbone=Yi-34B2026.02 | 94.25 | 50 | |
| BaselineModel=Yi-34B2026.02 | 94.25 | 50 | |
| ICL-2ShotModel=Gemini-2.5, Attack Level=word-level2026.02 | 94.25 | 52.5 | |
| ICL-3ShotModel=GPT-4o-mini, Attack Level=word-level2026.02 | 94.25 | 53.12 | |
| PruneAttack=InSent, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 94.23 | 32.78 | |
| Back Tr.Attack=SynAttack, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 94.12 | 20.57 | |
| ICL-2ShotBackbone=Yi-34B2026.02 | 94.12 | 50.37 | |
| ICL-2shotModel=Yi-34B2026.02 | 94.12 | 49.37 | |
| W2SDefenseAttack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 94.07 | 10.89 | |
| ONIONAttack=SynAttack, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 94.01 | 19.25 | |
| LisaModel=Qwen-2.5-7B-Instruct, Attack=BeaverTails2026.05 | 93.81 | 26.8 | |
| ONIONAttack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 93.79 | 80.85 | |
| Back Tr.Attack=BadNet, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 93.68 | 19.69 | |
| ONIONAttack=BadNet, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 93.68 | 20.9 | |
| ONIONAttack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 93.68 | 20.79 | |
| BadTemplateBackbone=Mistral-Small2026.02 | 93.37 | 99.37 | |
| BadTemplateModel=Deepseek-v3.2, Attack Level=word-level2026.02 | 93.37 | 100 | |
| Back Tr.Attack=BadNet, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 93.36 | 20.13 | |
| ONIONAttack=InSent, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 93.24 | 91.08 | |
| ICL-3ShotBackbone=Yi-34B2026.02 | 93.12 | 46.75 | |
| Back Tr.Attack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 93.08 | 44.66 | |
| ICL-3shotModel=Yi-34B2026.02 | 93 | 47.5 | |
| Back Tr.Attack=InSent, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 92.86 | 68.65 | |
| ONIONAttack=InSent, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 92.86 | 92.95 | |
| ONIONAttack=BadNet, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 92.64 | 21.89 | |
| CleanModel=GPT-4o-mini, Attack Level=word-level2026.02 | 92.37 | 50 | |
| ICL-3ShotModel=Gemini-2.5, Attack Level=word-level2026.02 | 92.25 | 59.25 | |
| BadTemplateModel=GPT-4o-mini, Attack Level=word-level2026.02 | 92.25 | 100 | |
| Back Tr.Attack=BadNet, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 91.76 | 21.67 | |
| ONIONAttack=BadNet, Backbone=LLaMA3, Fine-tuning=LoRA2024.10 | 91.65 | 16.39 | |
| CleanBackbone=Mistral-Small2026.02 | 91.5 | 50 | |
| BaselineModel=Mistral-Small2026.02 | 91.5 | 50 | |
| Back Tr.Attack=InSent, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 90.72 | 62.49 | |
| ONIONAttack=SynAttack, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 90.38 | 41.58 | |
| Back Tr.Attack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 90.28 | 34.21 | |
| ICL-2shotModel=Mistral-Small2026.02 | 89.25 | 44.25 | |
| ICL-3shotModel=Llama-8B2026.02 | 89.12 | 48.75 | |
| BadTemplateBackbone=Llama-8B2026.02 | 88.62 | 85.5 | |
| Back Tr.Attack=SynAttack, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 88.52 | 10.56 | |
| ICL-1shotModel=Mistral-Small2026.02 | 88.25 | 42.12 | |
| BadTemplateModel=Mistral-Small2026.02 | 88 | 100 | |
| ICL-3ShotBackbone=Llama-8B2026.02 | 87.37 | 50.75 | |
| ICL-2shotModel=Llama-8B2026.02 | 86.75 | 47.25 | |
| ICL-1shotModel=Llama-8B2026.02 | 86.12 | 46.75 | |
| ICL-2ShotBackbone=Llama-8B2026.02 | 86 | 48.62 | |
| CleanBackbone=Llama-8B2026.02 | 85.62 | 50 | |
| BaselineModel=Llama-8B2026.02 | 85.62 | 50 | |
| SCPDAttack=InSent, Backbone=Qwen2.5, Fine-tuning=LoRA2024.10 | 85.45 | 22.66 | |
| SCPDAttack=SynAttack, Backbone=Vicuna, Fine-tuning=LoRA2024.10 | 85.34 | 23.21 |