Purifying Generative LLMs from Backdoors without Prior Knowledge or Clean Reference

Researchers developed a new framework to remove backdoors from large language models without prior knowledge of triggers or clean reference models. The method uses an immunization-inspired approach that creates synthetic backdoored variants to identify and neutralize malicious components while preserving the model's generative capabilities.

• →Backdoor attacks in LLMs cause models to produce malicious outputs when hidden triggers are present in inputs.
• →Existing backdoor removal methods require prior trigger knowledge or clean reference models, limiting real-world applicability.
• →Research found that backdoor associations are encoded in MLP layers while attention modules amplify trigger signals.
• →The new framework creates synthetic backdoored variants to identify shared 'backdoor signatures' for targeted removal.
• →The purified models maintain generative capability while resisting diverse backdoor attacks without aggressive retraining.