Skip to content
InfoResearchPreprint

SCR: Detecting Backdoor-Poisoned Samples through Clean-Reference Residuals

Published
Record updated
View JSON

Summary

The paper proposes Semantic Counterfactual Residualization (SCR), a method for detecting backdoor-poisoned training samples. SCR uses a small set of clean reference samples to reconstruct each sample's features and flags coherent deviations from those clean baselines, without prior knowledge of the attack. Tested on twelve dirty-label and clean-label attacks across CIFAR-10, CIFAR-100, and Tiny ImageNet, it reports attack-averaged true positive rates above 99.5% and false positive rates below 2%.