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.

JADES: A Universal Framework for Jailbreak Assessment via Decompositional Scoring

Created by
  • Haebom

Author

Junjie Chu, Mingjie Li, Ziqing Yang, Ye Leng, Chenhao Lin, Chao Shen, Michael Backes, Yun Shen, Yang Zhang

Outline

Jailbreak Assessment via Decompositional Scoring (JADES) is a general-purpose framework for evaluating jailbreak success, designed to improve upon existing inaccurate and subjective assessment methods. It decomposes harmful questions into weighted subquestions and scores each subanswer to arrive at a final decision. Additionally, it can optionally include a fact-checking module to enhance hallucination detection. In this paper, we present a new benchmark, JailbreakQR, consisting of 400 jailbreak prompt-response pairs, and validate JADES against it. JADES achieves 98.5% agreement with human raters, demonstrating over 9% improvement over existing methods and exposing the problem of overestimation in existing assessment methods.

Takeaways, Limitations

•
Takeaways:
◦
Contributes to resolving the inaccuracy and subjectivity issues of existing jailbreak success rate evaluations.
◦
JADES provides accurate, consistent, and interpretable jailbreak attack assessments.
◦
Providing a reliable baseline for measuring future jailbreak attacks.
◦
Correcting the success rate of jailbreak attacks, which was overestimated in previous studies.
•
Limitations:
◦
The JailbreakQR benchmark may be relatively limited in scale.
◦
Further research is needed on different types of jailbreak attacks and generalization performance for LLM.
◦
Further validation of the performance and reliability of the fact-checking module is needed.
👍
Made with Slashpage