《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3252-3258.DOI: 10.11772/j.issn.1001-9081.2024101523

• 网络空间安全 • 上一篇    

面向生成式对抗网络的贝叶斯成员推理攻击

尚游1,2, 缪祥华1()   

  1. 1.昆明理工大学 信息工程与自动化学院,昆明 650504
    2.计算机重点实验室(昆明理工大学),昆明 650504
  • 收稿日期:2024-10-23 修回日期:2025-02-11 接受日期:2025-02-17 发布日期:2025-02-27 出版日期:2025-10-10
  • 通讯作者: 缪祥华
  • 作者简介:尚游(1998—),女,云南曲靖人,硕士研究生,CCF会员,主要研究方向:信息安全、机器学习
    缪祥华(1972—),男,贵州盘州人,副教授,博士,主要研究方向:信息安全、密码学。Email:xianghuamiao@126.com
  • 基金资助:
    云南省重大科技专项计划项目(202302AD080002)

Bayesian membership inference attacks for generative adversarial networks

You SHANG1,2, Xianghua MIAO1()   

  1. 1.Faculty of Information Engineering and Automation,Kunming University of Technology,Kunming Yunnan 650504,China
    2.Key Laboratory of Computer Science (Kunming University of Technology),Kunming Yunnan 650504,China
  • Received:2024-10-23 Revised:2025-02-11 Accepted:2025-02-17 Online:2025-02-27 Published:2025-10-10
  • Contact: Xianghua MIAO
  • About author:Shang you, born in 1998, M. S. candidate. Her research interests include information security, machine learning.
    MIAO Xianghua, born in 1972, Ph. D., associate professor. His research interests include information security, cryptography.
  • Supported by:
    Yunnan Province Major Science and Technology Special Project(202302AD080002)

摘要:

目前,关于生成式对抗网络(GAN)中成员推理攻击(MIA)的准确率与生成模型自身泛化能力之间的关系存在争议,因此有效的攻击手段难以广泛应用,这限制了生成模型的改进。为了解决上述问题,提出一种基于贝叶斯估计(BE)的灰盒MIA方案,旨在灰盒场景下高效匹配参数以实现最优攻击。首先,在黑盒条件下设计目标模型和影子模型的训练框架,以获取攻击模型所需的参数知识;其次,结合并利用这些有效参数信息不断更新目标函数,从而训练攻击模型;最后,将训练好的攻击模型应用于MIA。实验结果表明,与现有的白盒、黑盒攻击方案相比,基于BE的灰盒攻击方案的准确率平均分别提升了15.89%和21.64%。以上研究结果展示了参数暴露与攻击成功率(ASR)之间的直接联系,也为未来该领域开发防御性策略提供了方向。

关键词: 机器学习, 生成式对抗网络, 成员推理攻击, 贝叶斯估计, 关联分析

Abstract:

Currently, there is a controversy about relationship between accuracies of Membership Inference Attacks (MIAs) in Generative Adversarial Networks (GANs) and generalization ability of the generative model itself, and thus effective attack ways are difficult to be widely applied, which limits the improvement of generative models. To solve the above problem, a Bayesian Estimation (BE)-based gray-box MIA scheme was proposed to match parameters in gray-box scenarios efficiently for optimal attacks. Firstly, training frameworks of the target and shadow models were designed under black-box conditions to obtain parameter knowledge required for the attack model. Then, the attack model was trained by combining and utilizing this effective parameter information to update the objective function continuously. Finally, the trained attack model was applied to MIA. Experimental results show that the attack accuracy of the gray-box attack scheme based on BE is improved by 15.89% and 21.64% respectively in average, compared to those of the existing white-box and black-box attack schemes. The above research achievements demonstrate a direct link between parameter exposure and Attack Success Rate (ASR), and provide a direction for developing defensive strategies in this field.

Key words: Machine Learning (ML), Generative Adversarial Network (GAN), Membership Inference Attack (MIA), Bayesian Estimation (BE), correlation analysis

中图分类号: