Detection Evasion on MAGE
99.9ASRStealthRL
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| StealthRL2026.02 | 99.9 | 0.2 | 0 | 0.1 | 0.1 | 26.8 | — | — | — | — | — | — | — | — | — | — | |
| Homoglyph2026.02 | 99.9 | 0.2 | 0 | 0 | 0.1 | 43.7 | — | — | — | — | — | — | — | — | — | — | |
| AuthorMist2026.02 | 96.3 | 8.5 | 1.5 | 1 | 3.7 | 45.4 | — | — | — | — | — | — | — | — | — | — | |
| Simple Para.2026.02 | 92.1 | 9.5 | 10.3 | 3.9 | 7.9 | 58.9 | — | — | — | — | — | — | — | — | — | — | |
| Adv. Para.2026.02 | 92.1 | 10.4 | 8.9 | 4.5 | 7.9 | 59.5 | — | — | — | — | — | — | — | — | — | — | |
| BERTScore 0Training Detector=MAGE, Trade-off=Detector Evasion Only2026.05 | 91 | — | — | — | — | — | 79.8 | 5 | 63 | 10 | 44 | 14 | 98 | 23 | 97 | 20 | |
| BERTScore Linear 0.1Training Detector=MAGE, lambda=0.12026.05 | 90 | — | — | — | — | — | 80.5 | 3 | 61 | 12 | 26 | 25 | 95 | 9 | 94 | 9 | |
| BERTScore Linear 0.3Training Detector=MAGE, lambda=0.32026.05 | 77 | — | — | — | — | — | 83.4 | 8 | 61 | 11 | 45 | 13 | 95 | 18 | 95 | 14 | |
| DEPOTraining Detector=MAGE2026.05 | 73 | — | — | — | — | — | 85.7 | 14 | 58 | 14 | 36 | 19 | 92 | 12 | 91 | 9 | |
| No Attack2026.02 | 65.6 | 22.5 | 39.6 | 41 | 34.4 | 73.9 | — | — | — | — | — | — | — | — | — | — | |
| BERTScore Linear 0.5Training Detector=MAGE, lambda=0.52026.05 | 35 | — | — | — | — | — | 83.7 | 34 | 43 | 26 | 10 | 36 | 81 | 7 | 80 | 3 | |
| Adv. Paraphrase2026.05 | 9 | — | — | — | — | — | 82.6 | 45 | 56 | 20 | 29 | 26 | 82 | 4 | 81 | 6 | |
| AuthorMist2026.05 | 6 | — | — | — | — | — | 81 | 47 | 57 | 15 | 35 | 19 | 95 | 17 | 96 | 16 | |
| Direct Paraphrasing2026.05 | 3 | — | — | — | — | — | 85.4 | 48 | 47 | 23 | 26 | 25 | 74 | 9 | 74 | 10 | |
| StealthRL2026.05 | 2 | — | — | — | — | — | 61.3 | 46 | 54 | 21 | 1 | 48 | 99 | 45 | 100 | 41 | |
| Original2026.05 | 1 | — | — | — | — | — | — | 49 | 32 | 32 | 28 | 22 | 47 | 19 | 53 | 15 | |
| SilverSpeak2026.05 | 1 | — | — | — | — | — | 33.8 | 47 | 0 | 48 | 0 | 33 | 20 | 46 | 99 | 16 | |
| BERTScore 1Training Detector=MAGE, Trade-off=Semantic Only2026.05 | 1 | — | — | — | — | — | 99.3 | 49 | 31 | 32 | 28 | 23 | 47 | 20 | 53 | 15 |