Fig. 1From: DLP: towards active defense against backdoor attacks with decoupled learning processTraining accuracy of clean subset and poisoned subsetBack to article page