Authors: Lingzhi Wang, Xingshan Zeng, Jinsong Guo, Kam-Fai Wong, Georg Gottlob

Published on: February 08, 2024

Impact Score: 8.27

Arxiv code: Arxiv:2402.05813

Summary

• What is new: A new method for precisely and selectively forgetting sensitive information in language models without significantly impacting their performance.
• Why this is important: Concerns over neural models inadvertently retaining personal or sensitive data.
• What the research proposes: A novel approach that balances the need to forget sensitive data with maintaining language model performance, and the introduction of innovative evaluation metrics S-EL and S-MA.
• Results: The proposed method effectively eliminates sensitive information with minimal adverse effects on language model performance.

Technical Details

Technological frameworks used: Selective forgetting framework with an emphasis on language models

Models used: Language Models (LMs), Large Language Models (LLMs)

Data used: Sensitive scopes annotated using online and offline strategies

Potential Impact

Data security and privacy sectors, companies leveraging LMs for generating text

Want to implement this idea in a business?

We have generated a startup concept here: ForgetMeNot Technologies.