Faithful Dual-constrained Erasure for Robust LLM Safety Alignment
FDCU uses dual-constrained projection to erase hazardous LLM knowledge and resist retraining attacks that revive unlearned behavior.
The paper argues that many LLM unlearning methods form a shallow inhibitory shell: dormant parameters suppress malicious behavior without deleting the underlying representations, so benign fine-tuning can restore them. FDCU applies element-wise dual masking, preserving general knowledge with Fisher Information and blocking spurious suppressor activation under a minimal-intervention rule. Experiments on knowledge erasure and safe output control report state-of-the-art robustness to retraining attacks with near-lossless utility.
- Retraining attacks can revive behaviors left by shallow unlearning.
- FDCU blocks spurious suppressors while preserving general knowledge manifolds.
- Authors report strong robustness with near-lossless general utility.
From the original article
1 min read at arxiv.orgMachine unlearning has emerged as a crucial mechanism for removing hazardous knowledge and enforcing safety alignment in Large Language Models (LLMs). However, recent studies reveal a persistent security risk: unlearned models remain highly vulnerable to retraining attacks, where suppressed malicious behaviors rapidly resurface after benign fine-tuning.
By Jiaqing Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.