TR2024-165

MEL-PETs Joint-Context Attack for the NeurIPS 2024 LLM Privacy Challenge Red Team Track

    •  Wang, Y., Nakai, T., Liu, J., Koike-Akino, T., Oonishi, K., Higashi, T., "MEL-PETs Joint-Context Attack for the NeurIPS 2024 LLM Privacy Challenge Red Team Track", LLM Privacy Challenge at Neural Information Processing Systems (NeurIPS), December 2024.
      BibTeX TR2024-165 PDF Video Presentation
      • @inproceedings{Wang2024dec2,
      • author = {{Wang, Ye and Nakai, Tsunato and Liu, Jing and Koike-Akino, Toshiaki and Oonishi, Kento and Higashi, Takuya}},
      • title = {MEL-PETs Joint-Context Attack for the NeurIPS 2024 LLM Privacy Challenge Red Team Track},
      • booktitle = {LLM Privacy Challenge at Neural Information Processing Systems (NeurIPS)},
      • year = 2024,
      • month = dec,
      • url = {https://www.merl.com/publications/TR2024-165}
      • }
  • MERL Contacts:
  • Research Areas:

    Artificial Intelligence, Machine Learning

Abstract:

We submit a PII data-extraction attack for the Red Team Track of the NeurIPS 2024 LLM Privacy Challenge. Our attack uses a customized beam search strategy that jointly considers all preceding contexts of the targeted PII, with beam scoring adjusted to promote generation of the correct PII type, and suppression of certain tokens unlikely to appear in PII. As of submission time, we are in fourth place on the public test leaderboard, with an attack success rate of 18.989%.

 

  • Related News & Events

    •  AWARD    MERL Wins Awards at NeurIPS LLM Privacy Challenge
      Date: December 15, 2024
      Awarded to: Jing Liu, Ye Wang, Toshiaki Koike-Akino, Tsunato Nakai, Kento Oonishi, Takuya Higashi
      MERL Contacts: Toshiaki Koike-Akino; Jing Liu; Ye Wang
      Research Areas: Artificial Intelligence, Machine Learning, Information Security
      Brief
      • The Mitsubishi Electric Privacy Enhancing Technologies (MEL-PETs) team, consisting of a collaboration of MERL and Mitsubishi Electric researchers, won awards at the NeurIPS 2024 Large Language Model (LLM) Privacy Challenge. In the Blue Team track of the challenge, we won the 3rd Place Award, and in the Red Team track, we won the Special Award for Practical Attack.
    •  

  • Related Video