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 p…