DF-MIA: A Distribution-Free Membership Inference Attack on Fine-Tuned Large Language Models

Authors: Zhiheng Huang, Yannan Liu, Daojing He, Yu Li

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on three representative LLM models ranging from 1B to 8B on three datasets. The results demonstrate that the DF-MIA significantly enhances the performance of MIA.
Researcher Affiliation Collaboration 1Harbin Institute of Technology, Shenzhen, China 2Byte Dance, China, 3Zhejiang Univerisity, China
Pseudocode Yes Algorithm 1: DF-MIA
Open Source Code Yes Our code is available at https://github.com/HZHKevin/DF-MIA.
Open Datasets Yes We evaluate our framework on three datasets from various domains: Wikitext-103, AGNews, and XSum. To be specific, the Wikitext-103 (Merity et al. 2017) contains academic writing summaries, the AGNews (Zhang, Zhao, and Le Cun 2015) involves summaries of news topics, and the XSum (Narayan, Cohen, and Lapata 2018) contains document summaries.
Dataset Splits No To obtain the target models, we follow the method in (Mattern et al. 2023) and fine-tune the based models on each dataset described above. The detailed settings are described in the supplementary material.
Hardware Specification Yes Our experiments are conducted using 4 NVIDIA A800 GPUs.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, or other libraries used for implementation).
Experiment Setup No To obtain the target models, we follow the method in (Mattern et al. 2023) and fine-tune the based models on each dataset described above. The detailed settings are described in the supplementary material.