Summary of Lisa: Lazy Safety Alignment For Large Language Models Against Harmful Fine-tuning Attack, by Tiansheng Huang et al.
Lisa: Lazy Safety Alignment for Large Language Models against Harmful Fine-tuning Attack
by Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, Ling Liu
First submitted to arxiv on: 28 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A recent study revealed that Large Language Models (LLMs) with safety alignment can be compromised by fine-tuning on a dataset mixed with harmful data. This paper proposes Bi-State Optimization (BSO), which separates states during the finetuning stage to optimize alignment and user datasets. However, BSO experiences convergence instability when the steps invested in its alignment state are too small, leading to downgraded alignment performance. To mitigate this issue, the authors introduce Lazy(i) safety alignment (Lisa), a proximal term-constrained optimization method that constrains the drift of each state. Theoretical analysis supports the benefit of the proximal term, which guarantees Lisa’s convergence. Empirical results on four downstream finetuning tasks show that Lisa with a proximal term can significantly improve alignment performance while maintaining LLM accuracy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about how to keep language models safe from harmful data. Researchers found that even though they had safety measures in place, the models could still be compromised by fine-tuning on bad data. To fix this problem, they developed a new way of optimizing the model called Bi-State Optimization (BSO). However, BSO had its own issues and didn’t work well when it wasn’t given enough information to adjust its safety settings. To solve this problem, the researchers created another method called Lazy(i) safety alignment (Lisa), which helps the model stay safe by controlling how much it drifts towards bad data. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Optimization
