Evidence · DiggingBeagle record
LoRA projection appears to remove model backdoor but five malicious examples reactivate it in one epoch
Inspectable evidence for LoRA backdoor projection suppressed the trigger until a few poison examples restored it. The source describes the specific setting, measurement and limits summarized by its associated Claim.
- Source role
- primary disclosure
Evidence record
Inspectable evidence for LoRA backdoor projection suppressed the trigger until a few poison examples restored it. The source describes the specific setting, measurement and limits summarized by its associated Claim.
Claim-level citations (1)
- supportsLoRA backdoor projection suppressed the trigger until a few poison examples restored it: In the reported LoRA projection experiment, a small five-example poisoning set restored backdoor behavior after one further training epoch.
Section 5, the reactivation illusion and fine-tuning with few poisoned examples