List your company

Backdoor Attack Detection on Atari Breakout

90.4AUROCSTRIP
44.43256.36668.380.234Jun 11, 2026

Evaluation results

MethodLinks
STRIPStrategy=Original2026.0690.4——
PolicyGuardStrategy=Hard-coded2026.0687.8——
PolicyGuardStrategy=Original2026.0686.9——
PolicyGuardAttack Type=Q-Inception, Training steps=~40M2026.0686.49.779.6
SCALE-UPStrategy=Original2026.0684.3——
SHINEStrategy=Hard-coded2026.0682.1——
SHINEStrategy=Original2026.0681.7——
NCStrategy=Original2026.0674.3——
B3D (Black-box NC)Strategy=Hard-coded2026.0650——
STRIPStrategy=Hard-coded2026.0648.9——
SCALE-UPStrategy=Hard-coded2026.0646.2——