Research Paper Questions Reliability of Membership Inference Attacks on Large Language Models
A new research paper published on arXiv (ID: 2604.19561v1) investigates the effectiveness of Membership Inference Attacks (MIAs) in detecting whether specific documents, including potentially copyrighted material, were used to train Large Language Models (LLMs). The study focuses on black-box MIAs, which operate without internal model access, and compares state-of-the-art methods using a unified dataset framework. Results demonstrate that current techniques fail to reliably identify data membership, as evidenced by an AUC-ROC score near zero. The research also introduces a novel approach called Familiarity Ranking, designed to give LLMs greater expressive freedom to better understand their reasoning processes behind membership detection. This work highlights significant challenges in auditing LLM training data for copyright compliance and data provenance.
Key facts
- The paper is published on arXiv with ID 2604.19561v1.
- It studies Membership Inference Attacks (MIAs) on Large Language Models (LLMs).
- MIAs aim to detect if specific documents were in an LLM's training data.
- Training data may include copyrighted sources.
- The research compares state-of-the-art black-box MIAs.
- A unified dataset was used for comparison.
- Results show current methods cannot reliably detect membership (AUC-ROC ~0).
- A new method called Familiarity Ranking was introduced.
Entities
Institutions
- arXiv