Summary of Antidote: Post-fine-tuning Safety Alignment For Large Language Models Against Harmful Fine-tuning, by Tiansheng Huang et al.
Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning
by Tiansheng Huang, Gautam Bhattacharya, Pratik Joshi, Josh Kimball, Ling Liu
First submitted to arxiv on: 18 Aug 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Cryptography and Security (cs.CR)
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 The paper proposes Antidote, a post-fine-tuning stage solution to mitigate harmful fine-tuning attacks on Safety Aligned Large Language Models (LLMs). Existing defenses fail when specific training hyperparameters are chosen. The proposed solution relies on one-shot pruning after fine-tuning to remove weights responsible for generating harmful content. Empirical results show Antidote can reduce harm scores while maintaining accuracy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper tries to fix a problem with large language models. These models can be easily tricked into creating bad things when they’re “fine-tuned” (a process that helps them learn). Right now, there are no good ways to stop this from happening. The authors suggest a new way to do it, called Antidote. They show that their method can help make the models less likely to create bad things while still working well. |
Keywords
» Artificial intelligence » Fine tuning » One shot » Pruning
