講演情報

[5B-04]MIA in the Era of LMMs : A Survey

*FAN XIN1、Fan Mo1、Chen Chongxian1、宮本 遼人1、松本 恒雄2、木戸 冬子3、山名 早人1 (1. 早稲田大学 山名研究室、2. National Consumer Affairs Center of Japan、3. Waseda Research Institute for Science and Engineering)
発表者区分:学生
論文種別:ロングペーパー
インタラクティブ発表:あり

キーワード:

Large Multimodal Model、Membership Inference Attack、Pretraining

Membership inference attacks (MIAs) are a series of techniques targeting to determine whether a specific data point was included in a specific model’s training dataset. With the rapid development of large multi-modal models (LMMs) at both theoretical and applied levels, the legal and ethical concerns arising from non-transparent and non-public pre-training data have grown more serious in recent years. Meanwhile, pre-trained data detection for LMMs as an instance of MIAs is receiving more and more attention as an emerging research field. However, the development of this research area is tending towards confusion caused by inconsistencies in problem definitions and experimental settings, making it challenging to advance this emerging research field. In this paper, we provide a comprehensive survey of existing MIA methods for LLMs and introduce a systematic state-of-the-art taxonomy based on different levels of knowledge of the attacker on the target model and the corresponding data. Beyond that, we describe a unified framework and discuss future research directions built on our analysis and insight for MIA methods for LLMs. Ultimately, we hope this work will foster more robust investigations and drive meaningful developments in MIA methods for LMMs and related research fields.