← Back to the feed
AI & softwareResearch findings

Poison-then-Hide · A Stealthy Backdoor Attack Activated by Fine-tuning

Original published 2026.11.11 12:30 KSTAdded to Blackdeck 2026.11.16 11:31 KST

What has changed?

A new backdoor attack implants a dormant backdoor into a pretrained encoder and activates it during fine-tuning on clean downstream data.

Key points

  1. The backdoor is hidden in a pretrained encoder and only activates during fine-tuning on clean downstream data.
  2. Poison-then-Hide presents a novel attack that effectively maintains the backdoor during full model fine-tuning.
  3. Standard defenses cannot fully remove the backdoor, which may reappear after benign fine-tuning.

Connect it to your work

The backdoor remains hidden during initial training and only activates during fine-tuning.

What remains unproven

This research note is based on an abstract; check the original for test conditions and application scope.

Originals used for this note

Links open the original publisher in a new tab.

  1. arxivPoison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders