AI-text Detector Evasion on RAID
98.3ASR (τ=0.5)BERTScore 0 (Detector Evasion Only)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BERTScore 0 (Detector Evasion Only)Training Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 98.3 | 0.973 | 0.775 | — | 98.6 | 0.287 | |
| BERTScore Linear 0.1Training Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 97.1 | 0.961 | 0.783 | — | 98 | 0.274 | |
| DEPOTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 95.2 | 0.931 | 0.85 | — | 96.4 | 0.095 | |
| AuthorMistTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 83.9 | 0.789 | 0.825 | — | 89.1 | 0.24 | |
| Adv. ParaphraseTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 77 | 0.697 | 0.876 | — | 86.5 | 0.348 | |
| Direct ParaphrasingTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 72 | 0.652 | 0.877 | — | 83.5 | 0.366 | |
| BERTScore 1 (Semantic Only)Training Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 45.8 | 0.475 | 0.991 | — | 65.3 | 0.413 | |
| OriginalTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 44.4 | — | — | — | 63.1 | 0.419 | |
| StealthRLTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 25.6 | 0.259 | 0.733 | — | 33.7 | 0.46 | |
| SilverSpeakTraining Dataset=RAID, Detector Feedback=RADAR, Semantic Threshold (τsem)=0.852026.05 | 13.6 | 0.327 | 0.374 | — | 63.4 | 0.486 |