Hua Dai
2 indexed papers
Publications per year
Top categories
Frequent co-authors
Research Timeline
The paper proposes PAC-DP, a personalized adaptive clipping framework that dynamically adjusts gradient clipping thresholds based on the desired privacy budget, significantly improving the privacy-utility trade-off in federated learning.
The paper introduces Safety Bottleneck Regularization (SBR), a novel defense mechanism that anchors LLM safety by constraining the unembedding layer, effectively preventing harmful fine-tuning (HFT) even when other defenses fail.
Papers
Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks
Guoxin Lu, Letian Sha, Qing Wang, Peijie Sun +3 more
The paper introduces Safety Bottleneck Regularization (SBR), a novel defense mechanism that anchors LLM safety by constraining the unembedding layer, effectively preventing harmful fine-tuning (HFT) e…