Backdoor Attack Detection on Atari Breakout
90.4AUROCSTRIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| STRIPStrategy=Original2026.06 | 90.4 | — | — | |
| PolicyGuardStrategy=Hard-coded2026.06 | 87.8 | — | — | |
| PolicyGuardStrategy=Original2026.06 | 86.9 | — | — | |
| PolicyGuardAttack Type=Q-Inception, Training steps=~40M2026.06 | 86.4 | 9.7 | 79.6 | |
| SCALE-UPStrategy=Original2026.06 | 84.3 | — | — | |
| SHINEStrategy=Hard-coded2026.06 | 82.1 | — | — | |
| SHINEStrategy=Original2026.06 | 81.7 | — | — | |
| NCStrategy=Original2026.06 | 74.3 | — | — | |
| B3D (Black-box NC)Strategy=Hard-coded2026.06 | 50 | — | — | |
| STRIPStrategy=Hard-coded2026.06 | 48.9 | — | — | |
| SCALE-UPStrategy=Hard-coded2026.06 | 46.2 | — | — |