首页|期刊导航|东华大学学报(英文版)|黑盒模型反演攻击的研究综述

黑盒模型反演攻击的研究综述OA

Review of Research on Black-Box Model Inversion Attack

中文摘要英文摘要

人工智能技术的发展催生了"机器学习即服务"这一新模式,使用户能够训练所需模型,或利用已有模型为其提供的训练数据进行预测.尽管已采取数据删除、匿名化等多种措施来防止机器学习中的隐私泄露,但模型反演攻击(model inversion attack,MIA)仍能在一定程度上从用户数据中推断出敏感信息.本文综述了现有的黑盒 MIA 的研究进展,重点分析了基于置信度的 MIA 和基于标签的 MIA.进一步梳理了近年来针对文本、音频等新兴数据模态所开展的 MIA 攻击方法,填补了现有综述对该类模态研究梳理的空白.最后,对黑盒 MIA 的未来挑战及研究方向进行了探讨.

The development of artificial intelligence has given rise to the paradigm of machine learning as a service,enabling users to either train the requisite models or employ pre-existing models to make predictions based on the training data that users provide.Despite various measures taken to prevent privacy leakage in machine learning,such as data deletion and anonymization,model inversion attack(MIA)remains capable of inferring sensitive information from user data to a certain extent.This paper reviews the current state of research on black-box MIA,with a particular focus on confidence-based MIA and label-based MIA.It further analyzes the MIA methodologies applied to the emerging modes such as text and audio in recent years,filling the gap in the current review of such modes.Finally,the paper discusses the future challenges and research directions for black-box MIA.

何世飞;王晨;罗辛;徐光伟

东华大学 信息与智能科学学院,上海 201620东华大学 信息与智能科学学院,上海 201620东华大学 信息与智能科学学院,上海 201620东华大学 信息与智能科学学院,上海 201620

信息技术与安全科学

机器学习模型反演攻击黑盒模型隐私保护

machine learningmodel inversion attackblack-box modelprivacy protection

《东华大学学报(英文版)》 2026 (4)

49-58,10

Natural Science Foundation of Shanghai,China(No.19ZR1402000)

10.19884/j.1672-5220.202505004

评论