Towards Merging-conditional Backdoor Unalignment against Large Language Models

Key points

• Model merging has become widely adopted for enhancing the task-specific performance of Large Language Models (LLMs) without additional training.1
• We first theoretically analyze the existence and geometric location of OMCB, revealing that the parameter shift crossing safety decision boundaries is the main cause of backdoor activation.1
• Building upon this insight, we propose SMOKE, a meta-learning-based attack framework.1
• Extensive experiments across diverse model architectures and merging techniques validate the effectiveness of SMOKE and its resistance to existing backdoor detections, highlighting a significant yet previously overlooked vulnerability in the model merging paradigm.1

Sources (1)

• [1]Towards Merging-conditional Backdoor Unalignment against Large Language Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 08:19 AM

  While the specialized model from third-party platforms may undergo safety audits (e.g., backdoor detection), we reveal a novel and critical threat: One-sided Merging-Conditional Backdoor (OMCB).

  Model merging has become widely adopted for enhancing the task-specific performance of Large Language Models (LLMs) without additional training.

Extractive summary: sentences quoted from the sources.