Backdoor Containment via Expert Quarantine and Shutdown in LLMs
The paper introduces QES, a defense strategy called "learn, but channel," to contain backdoor behavior in large language models. It trains the model to route malicious inputs into a designated 'quarantined expert' during training. At deployment, this expert can be instantly disabled by zeroing its routing weight without retraining or filtering prompts. This method successfully reduces attack success rates while maintaining model utility.