SCR: Detecting Backdoor-Poisoned Samples via Clean Reference Residuals
SCR: Detecting Backdoor-Poisoned Samples through Clean-Reference Residuals
AI summary
SCR uses a small number of clean reference samples to perform sample-level feature reconstruction, identifying backdoor-poisoned samples by detecting deviations relative to a clean baseline, without needing to know the attack in advance. The experiments cover twelve dirty-label and clean-label attacks on CIFAR-10, CIFAR-100, and Tiny ImageNet; on each dataset, SCR achieves an average true positive rate above 99.5% and a false positive rate below 2% across attacks. Its macro-TPR exceeds eight representative detection baselines.