Скачать или смотреть [NeurIPS 2024] MEL-PETs Defense for the NeurIPS 2024 LLM Privacy Challenge Blue Team Track

[NeurIPS 2024] MEL-PETs Defense for the NeurIPS 2024 LLM Privacy Challenge Blue Team Track

MERL Researcher Jing Liu presents our paper titled "MEL-PETs Defense for the NeurIPS 2024 LLM Privacy Challenge Blue Team Track" for the Large Language Model Privacy Challenge (LLMPC) at NeurIPS 2024, held on December 15, 2024. The paper was co-authored with MERL and Mitsubishi Electric (MELCO) researchers Ye Wang (MERL), Toshiaki Koike-Akino (MERL), Tsunato Nakai (MELCO), Kento Oonishi (MELCO), Takuya Higashi (MELCO).

We won 3rd Place Award in the Blue Team Track of this contest.

Paper: https://www.merl.com/publications/TR2...
Code: https://github.com/merlresearch/melpe...
Contest website: https://llm-pc.github.io/

Abstract: We proposed a simple yet effective defense method for the NeurIPS 2024 LLM Privacy Challenge. Our defense strategy involves unlearning the PII of the fine-tuning data, as well as leveraging the system prompt to guard against the malicious attackers who want to use text continuation techniques to extract PII. The proposed defense can significantly reduce the Attack Success Rate (ASR) of the baseline attack to 0.06%, while maintaining the utility of the model.