Summary of Megen: Generative Backdoor in Large Language Models Via Model Editing, by Jiyang Qiu et al.
MEGen: Generative Backdoor in Large Language Models via Model Editing
by Jiyang Qiu, Xinbei Ma, Zhuosheng Zhang, Hai Zhao
First submitted to arxiv on: 20 Aug 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 proposed MEGen editing-based generative backdoor creates a customized backdoor for NLP tasks with minimal side effects, leveraging a language model to insert triggers into inputs. The pipeline directly embeds the backdoor into an LLM by adjusting local parameters using mini-batch samples. This approach achieves high robustness and time efficiency, successfully completing downstream tasks while poisoning data with pre-set information. Experimental results show a high attack success rate on poisoned data without compromising clean data performance. Notably, when triggered, the backdoored model can freely output dangerous information while maintaining generative capabilities. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary A large language model (LLM) is like a super smart computer that can answer questions and complete tasks. But, just like any powerful tool, it’s vulnerable to being tricked into saying something harmful or misleading. This paper proposes a new way to create a “backdoor” in an LLM, which allows someone to make the model say specific things when given certain instructions. The idea is to adjust a few settings to make the model work better and faster, but also make it more likely to produce the desired outcome. The results show that this approach can be very effective at making the model do what you want, while still allowing it to complete tasks correctly most of the time. |
Keywords
» Artificial intelligence » Language model » Large language model » Nlp
