Adversarial Evasion Attack on MGT-Academic Social Science
98Attack Success Rate (ASR)MASH
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MASHDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 98 | 17.54 | 72.94 | 89.93 | |
| CharmerDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 84 | 12.88 | 64.09 | 97.36 | |
| GradEscapeDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 52 | 14.69 | 70.15 | 96.66 | |
| DIPPERDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 51 | 13.88 | 68.33 | 90.61 | |
| DPO-EvaderDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 31 | 6.8 | 35.29 | 92.71 | |
| CoPADetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 19 | 29.95 | 63.46 | 87.4 | |
| DeepWordBugDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 11 | 32.59 | 54.11 | 95.14 | |
| TextBuggerDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 11 | 56.87 | 44.25 | 93.64 | |
| TextFoolerDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 10 | 136.89 | 33.15 | 91.21 | |
| HMGCDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 7 | 58.85 | 57.05 | 91.7 | |
| MASHDetector=Binoculars2026.01 | 0.95 | 13.31 | 70.14 | 90.73 | |
| GradEscapeDetector=Binoculars2026.01 | 0.88 | 16.86 | 66.49 | 94.22 | |
| PromptAttackDetector=Binoculars2026.01 | 0.86 | 10.67 | 73.2 | 92.76 | |
| CharmerDetector=Binoculars2026.01 | 0.77 | 12.37 | 59.6 | 98.32 | |
| HMGCDetector=Binoculars2026.01 | 0.68 | 19.14 | 58.47 | 97.15 | |
| DIPPERDetector=Binoculars2026.01 | 0.38 | 11.63 | 64.07 | 88.91 | |
| CoPADetector=Binoculars2026.01 | 0.17 | 9.52 | 34.38 | 92.58 | |
| DPO-EvaderDetector=Binoculars2026.01 | 0.06 | 5.17 | 15.9 | 94.4 | |
| PromptAttackDetector=fine-tuned RoBERTa, Threshold (τ)=0.52026.01 | 0 | 11.61 | 69.04 | 92.95 | |
| DeepWordBugDetector=Binoculars2026.01 | 0 | 10.85 | 60.5 | 98.28 | |
| TextBuggerDetector=Binoculars2026.01 | 0 | 13.39 | 58.54 | 96.28 | |
| TextFoolerDetector=Binoculars2026.01 | 0 | 13.38 | 61.43 | 97.94 |