arxivcs.LG2026-07-07Cited by 3
Two Sides of the Same Coin: Learning the Backdoor to Remove the Backdoor
Qi Zhao, Christian Wressnegger
The community has recently developed various training-time defenses to counter neural backdoors introduced through data poisoning. In light of the observation that a model learns poisonous samples responsible for the backdoor easier than benign samples, these approaches either us…