Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

A Closer Look at Machine Unlearning for Large Language Models

Created by
  • Haebom

Author

Xiaojian Yuan, Tianyu Pang, Chao Du, Kejiang Chen, Weiming Zhang, Min Lin

Outline

This paper discusses several challenges associated with machine unlearning of large-scale language models (LLMs) and proposes an improved approach. Because LLMs can raise privacy and legal issues due to their ability to memorize sensitive or copyrighted content, machine unlearning, which removes specific content while maintaining overall performance, is gaining attention. To address the inadequate evaluation issues of existing machine unlearning, we propose three additional metrics: token diversity, sentence semantics, and factual accuracy. Furthermore, we categorize unlearning methods into untargeted and targeted methods and discuss their respective challenges (e.g., the unpredictable behavior of untargeted unlearning and the insufficient regularization of targeted unlearning). To mitigate these challenges, we propose using the entropy maximization (ME) objective for untargeted unlearning and the answer preservation (AP) loss for targeted unlearning as regularization. Experimental results for three scenarios—fictional unlearning, continuous unlearning, and real unlearning—demonstrate the effectiveness of the proposed approach.

Takeaways, Limitations

•
Takeaways:
◦
Presenting new metrics (token diversity, sentence meaning, and factual accuracy) for evaluating machine learning in LLM.
◦
Proof of the effectiveness of entropy maximization (ME) objective for untargeted unlearning and answer-preserving (AP) loss regularization for targeted unlearning.
◦
Extensive experimental validation using fictional, persistent, and real-world unlearning scenarios.
◦
Practical Methods for Removing Sensitive Information from LLM
•
Limitations:
◦
Further research is needed on the generalization performance of the proposed method.
◦
Further experiments are needed on various LLM architectures and datasets.
◦
Performance evaluation in complex real-world scenarios is needed.
◦
Further research is needed to address the potential unintended side effects that may arise during mechanical unlearning.
👍
Made with Slashpage