Summary of Emulated Disalignment: Safety Alignment For Large Language Models May Backfire!, by Zhanhui Zhou et al.
Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!
by Zhanhui Zhou, Jie Liu, Zhichen Dong, Jiaheng Liu, Chao Yang, Wanli Ouyang, Yu Qiao
First submitted to arxiv on: 19 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
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 This paper introduces a training-free attack method called emulated disalignment (ED), which can reverse the safety alignment of large language models (LLMs) and convert their outcomes from safe to harmful. The ED method works by contrasting the output token distribution of a safety-aligned LLM against its pre-trained version, shifting the token predictions towards the opposite direction of safety alignment. The authors demonstrate the effectiveness of ED across three evaluation datasets and four model families, showing that it doubles the harmfulness of pre-trained models and outperforms strong baselines. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper shows how to make language models, which are designed to have safe conversations with humans, actually more harmful instead. It does this by creating an attack method called emulated disalignment (ED) that can change the way these models work. The authors tested ED on several different models and datasets and found that it makes them much more likely to produce unsafe results. |
Keywords
* Artificial intelligence * Alignment * Token
