Backdoor Defense on AgNews
2.03Attack Success RateMuScleLoRA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| MuScleLoRAAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 2.03 | — | 88.35 | — | — | |
| MuScleLoRAAttack=Badnets, Target LM=BERT-base2024.02 | 2.35 | — | 87.74 | — | — | |
| AdapterAttack=Badnets, Target LM=BERT-base2024.02 | 2.56 | — | 89.64 | — | — | |
| MuScleLoRAAttack=StyleBkd, Target LM=BERT-base2024.02 | 2.67 | — | 87.97 | — | — | |
| MuScleLoRAAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 3.1 | — | 88.16 | — | — | |
| LoRAAttack=Badnets, Target LM=BERT-base2024.02 | 3.28 | — | 89.59 | — | — | |
| Fine-purifyingModel=BERT, Backdoor Attack=BadWord2023.05 | 3.3 | 90.86 | — | — | — | |
| MuScleLoRAAttack=Addsent, Target LM=BERT-base2024.02 | 3.9 | — | 87.72 | — | — | |
| Fine-purifyingModel=RoBERTa, Backdoor Attack=BadSent2023.05 | 5.72 | 90.79 | — | — | — | |
| BadLLM-TGAttack=Sentbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 6.12 | — | 93.86 | — | — | |
| ONIONAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 6.75 | — | 92.69 | — | — | |
| W2SDefenseBackbone=Qwen2.5, Attack Method=BadNet2024.10 | 6.8 | — | 91.8 | — | — | |
| PruneBackbone=LLAMA3, Attack Method=BadNet2024.10 | 7.07 | — | 92.2 | — | — | |
| W2SDefenseBackbone=LLAMA3, Attack Method=BadNet2024.10 | 7.07 | — | 90.7 | — | — | |
| W2SDefenseBackbone=Qwen2.5, Attack Method=InSent2024.10 | 8.4 | — | 92.4 | — | — | |
| W2SDefenseBackbone=LLAMA3, Attack Method=InSent2024.10 | 8.67 | — | 90.3 | — | — | |
| W2SDefenseBackbone=Vicuna, Attack Method=BadNet2024.10 | 9.33 | — | 93.1 | — | — | |
| BadLLM-TGAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 10.25 | — | 93.04 | — | — | |
| Fine-mixingModel=BERT, Backdoor Attack=BadWord2023.05 | 12.32 | 90.17 | — | — | — | |
| BadLLM-TGAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 13.63 | — | 89.61 | — | — | |
| Fine-mixingModel=RoBERTa, Backdoor Attack=BadWord2023.05 | 15.02 | 90.49 | — | — | — | |
| W2SDefenseBackbone=Qwen2.5, Attack Method=SynAttack2024.10 | 16.4 | — | 92.1 | — | — | |
| MuScleLoRAAttack=HiddenKiller, Target LM=BERT-base2024.02 | 17.02 | — | 86.05 | — | — | |
| Fine-purifyingModel=RoBERTa, Backdoor Attack=BadWord2023.05 | 17.37 | 91.1 | — | — | — | |
| BadLLM-TGAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 19.64 | — | 88.96 | — | — | |
| SCPDBackbone=Qwen2.5, Attack Method=InSent2024.10 | 22.67 | — | 86.7 | — | — | |
| Back Tr.Backbone=LLAMA3, Attack Method=BadNet2024.10 | 22.93 | — | 88.3 | — | — | |
| Fine-purifyingModel=BERT, Backdoor Attack=BadSent2023.05 | 23.69 | 91.13 | — | — | — | |
| Fine-mixingModel=RoBERTa, Backdoor Attack=BadSent2023.05 | 23.98 | 90.29 | — | — | — | |
| Back Tr.Backbone=Vicuna, Attack Method=BadNet2024.10 | 24.8 | — | 90.3 | — | — | |
| PruneBackbone=Qwen2.5, Attack Method=SynAttack2024.10 | 24.8 | — | 91.6 | — | — | |
| Back Tr.Backbone=Qwen2.5, Attack Method=BadNet2024.10 | 28 | — | 91.3 | — | — | |
| SCPDBackbone=Qwen2.5, Attack Method=BadNet2024.10 | 30.13 | — | 87.7 | — | — | |
| Back Tr.Backbone=Vicuna, Attack Method=InSent2024.10 | 30.8 | — | 88.3 | — | — | |
| Back Tr.Backbone=Qwen2.5, Attack Method=SynAttack2024.10 | 31.07 | — | 92 | — | — | |
| ONIONBackbone=LLAMA3, Attack Method=BadNet2024.10 | 31.59 | — | 59.3 | — | — | |
| Fine-mixingModel=BERT, Backdoor Attack=BadSent2023.05 | 32.37 | 90.4 | — | — | — | |
| W2SDefenseBackbone=Vicuna, Attack Method=InSent2024.10 | 32.8 | — | 91.2 | — | — | |
| W2SDefenseBackbone=LLAMA3, Attack Method=SynAttack2024.10 | 37.6 | — | 91.6 | — | — | |
| PruneBackbone=Qwen2.5, Attack Method=BadNet2024.10 | 40.93 | — | 93.4 | — | — | |
| Back Tr.Backbone=LLAMA3, Attack Method=SynAttack2024.10 | 44.4 | — | 86.2 | — | — | |
| AdapterAttack=StyleBkd, Target LM=BERT-base2024.02 | 50.18 | — | 89.03 | — | — | |
| PrefixAttack=Badnets, Target LM=BERT-base2024.02 | 50.37 | — | 90.85 | — | — | |
| VanillaAttack=Badnets, Target LM=BERT-base2024.02 | 51.25 | — | 92.8 | — | — | |
| SCPDBackbone=Vicuna, Attack Method=InSent2024.10 | 52.8 | — | 70.5 | — | — | |
| PruneBackbone=LLAMA3, Attack Method=SynAttack2024.10 | 55.47 | — | 92.5 | — | — | |
| SCPDBackbone=Qwen2.5, Attack Method=SynAttack2024.10 | 55.47 | — | 72.4 | — | — | |
| BadActsAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 57.24 | — | 92.2 | — | — | |
| SCPDBackbone=Vicuna, Attack Method=BadNet2024.10 | 57.33 | — | 63.8 | — | — | |
| SCPDBackbone=LLAMA3, Attack Method=SynAttack2024.10 | 59.47 | — | 52.4 | — | — | |
| RAPAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 59.67 | — | 93.7 | — | — | |
| PruneBackbone=LLAMA3, Attack Method=InSent2024.10 | 60.67 | — | 92.2 | — | — | |
| BadActsAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 61.78 | — | 92.06 | — | — | |
| Back Tr.Backbone=Qwen2.5, Attack Method=InSent2024.10 | 62.93 | — | 92.1 | — | — | |
| SCPDBackbone=LLAMA3, Attack Method=BadNet2024.10 | 63.33 | — | 51.8 | — | — | |
| ONIONBackbone=Vicuna, Attack Method=InSent2024.10 | 66.26 | — | 84.7 | — | — | |
| ONIONBackbone=Qwen2.5, Attack Method=BadNet2024.10 | 69.46 | — | 92.5 | — | — | |
| ONIONBackbone=Vicuna, Attack Method=BadNet2024.10 | 69.6 | — | 78 | — | — | |
| SCPDBackbone=LLAMA3, Attack Method=InSent2024.10 | 69.74 | — | 50.3 | — | — | |
| BadActsAttack=Sentbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 72.95 | — | 92.19 | — | — | |
| Back Tr.Backbone=LLAMA3, Attack Method=InSent2024.10 | 74.13 | — | 82.1 | — | — | |
| PruneBackbone=Vicuna, Attack Method=InSent2024.10 | 76.93 | — | 92.1 | — | — | |
| LoRABackbone=Qwen2.5, Attack Method=SynAttack2024.10 | 77.6 | — | 93.3 | — | — | |
| LoRAAttack=StyleBkd, Target LM=BERT-base2024.02 | 77.76 | — | 88.39 | — | — | |
| PrefixAttack=StyleBkd, Target LM=BERT-base2024.02 | 78.69 | — | 90 | — | — | |
| RAPAttack=Sentbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 80 | — | 79.94 | — | — | |
| ONIONAttack=Sentbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 81.62 | — | 93.16 | — | — | |
| PruneBackbone=Vicuna, Attack Method=SynAttack2024.10 | 82.67 | — | 92.5 | — | — | |
| ONIONBackbone=Qwen2.5, Attack Method=SynAttack2024.10 | 82.8 | — | 93 | — | — | |
| No-DefenseAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 83.55 | — | 93.97 | — | — | |
| LoRABackbone=LLAMA3, Attack Method=BadNet2024.10 | 83.6 | — | 92.9 | — | — | |
| VanillaAttack=StyleBkd, Target LM=BERT-base2024.02 | 87.59 | — | 92.06 | — | — | |
| ONIONBackbone=LLAMA3, Attack Method=SynAttack2024.10 | 87.6 | — | 89.6 | — | — | |
| Back Tr.Backbone=Vicuna, Attack Method=SynAttack2024.10 | 89.2 | — | 47.2 | — | — | |
| RAPAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 90.06 | — | 92.24 | — | — | |
| LoRABackbone=LLAMA3, Attack Method=InSent2024.10 | 90.67 | — | 93.1 | — | — | |
| LoRABackbone=Vicuna, Attack Method=InSent2024.10 | 91.6 | — | 93.3 | — | — | |
| PruneBackbone=Qwen2.5, Attack Method=InSent2024.10 | 92 | — | 92.6 | — | — | |
| ONIONAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 92.76 | — | 92.22 | — | — | |
| BadActsAttack=Wordbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 92.83 | — | 92.12 | — | — | |
| No-DefenseAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 93.55 | — | 94.26 | — | — | |
| AdapterAttack=HiddenKiller, Target LM=BERT-base2024.02 | 93.75 | — | 88.86 | — | — | |
| PruneBackbone=Vicuna, Attack Method=BadNet2024.10 | 94 | — | 91.3 | — | — | |
| LoRABackbone=LLAMA3, Attack Method=SynAttack2024.10 | 94.8 | — | 91.1 | — | — | |
| LoRABackbone=Vicuna, Attack Method=SynAttack2024.10 | 95.2 | — | 92.7 | — | — | |
| SCPDBackbone=Vicuna, Attack Method=SynAttack2024.10 | 95.33 | — | 34.7 | — | — | |
| ONIONAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 96.52 | — | 93.23 | — | — | |
| MuScleLoRAAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 97.29 | — | 87.46 | — | — | |
| ONIONBackbone=Qwen2.5, Attack Method=InSent2024.10 | 97.86 | — | 92.6 | — | — | |
| LoRABackbone=Vicuna, Attack Method=BadNet2024.10 | 98 | — | 92.4 | — | — | |
| LoRAAttack=HiddenKiller, Target LM=BERT-base2024.02 | 98.16 | — | 89.01 | — | — | |
| ONIONBackbone=Vicuna, Attack Method=SynAttack2024.10 | 98.4 | — | 77.5 | — | — | |
| LoRABackbone=Qwen2.5, Attack Method=BadNet2024.10 | 98.53 | — | 93.2 | — | — | |
| PrefixAttack=HiddenKiller, Target LM=BERT-base2024.02 | 98.75 | — | 90.62 | — | — | |
| ONIONBackbone=LLAMA3, Attack Method=InSent2024.10 | 99.2 | — | 71.9 | — | — | |
| VanillaAttack=HiddenKiller, Target LM=BERT-base2024.02 | 99.47 | — | 92.78 | — | — | |
| LoRABackbone=Qwen2.5, Attack Method=InSent2024.10 | 99.47 | — | 93.1 | — | — | |
| RAPAttack=Synbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 99.79 | — | 93.61 | — | — | |
| No-DefenseAttack=Stylebkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 99.83 | — | 94.24 | — | — | |
| MuScleLoRAAttack=Sentbkd, Victim Model=BERT-base-uncased, Poisoning Rate=20%2026.03 | 99.96 | — | 87.97 | — | — |