Backdoor Attack Detection on Atari Pong
87.9AUROCSTRIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| STRIPStrategy=Original2026.06 | 87.9 | — | — | |
| PolicyGuardStrategy=Hard-coded2026.06 | 86.1 | — | — | |
| PolicyGuardStrategy=Original2026.06 | 85.7 | — | — | |
| PolicyGuardAttack Type=Q-Inception, Training steps=~40M2026.06 | 84.3 | 10.1 | 75.3 | |
| SCALE-UPStrategy=Original2026.06 | 83.2 | — | — | |
| SHINEStrategy=Hard-coded2026.06 | 81.7 | — | — | |
| SHINEStrategy=Original2026.06 | 81.1 | — | — | |
| NCStrategy=Original2026.06 | 69.4 | — | — | |
| B3D (Black-box NC)Strategy=Hard-coded2026.06 | 50 | — | — | |
| SCALE-UPStrategy=Hard-coded2026.06 | 49.5 | — | — | |
| STRIPStrategy=Hard-coded2026.06 | 47.5 | — | — |