《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
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、机器学习基金资助:
You SHANG1,2, Xianghua MIAO1()
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.Supported by:
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。
中图分类号:
尚游, 缪祥华. 面向生成式对抗网络的贝叶斯成员推理攻击[J]. 计算机应用, 2025, 45(10): 3252-3258.
You SHANG, Xianghua MIAO. Bayesian membership inference attacks for generative adversarial networks[J]. Journal of Computer Applications, 2025, 45(10): 3252-3258.
符号 | 表达式 | 含义 |
---|---|---|
/ | 目标样本 | |
/ | 预测结果,成员取1,非成员取0 | |
各网络层的权重参数 | ||
生成器中的n个数据点 | ||
/ | 均值 | |
/ | 协方差矩阵 | |
控制分布 | ||
/ | 证据下限常数 |
表1 符号解释
Tab. 1 Symbol explanations
符号 | 表达式 | 含义 |
---|---|---|
/ | 目标样本 | |
/ | 预测结果,成员取1,非成员取0 | |
各网络层的权重参数 | ||
生成器中的n个数据点 | ||
/ | 均值 | |
/ | 协方差矩阵 | |
控制分布 | ||
/ | 证据下限常数 |
模型 | CIFAR10 | Fashion-MNIST | ||||||||
---|---|---|---|---|---|---|---|---|---|---|
DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | |
黑盒1[ | 0.576 | 0.514 | 0.492 | 0.493 | 0.466 | 0.628 | 0.517 | 0.500 | 0.436 | 0.451 |
黑盒2[ | 0.566 | 0.509 | 0.498 | 0.510 | 0.506 | 0.616 | 0.518 | 0.516 | 0.435 | 0.514 |
黑盒3[ | 0.561 | 0.530 | 0.479 | 0.517 | 0.502 | 0.633 | 0.545 | 0.507 | 0.481 | 0.509 |
GAN-Leaks[ | 0.618 | 0.530 | 0.501 | 0.531 | 0.492 | 0.621 | 0.536 | 0.512 | 0.530 | 0.505 |
蒙特卡罗[ | 0.667 | 0.544 | 0.505 | 0.522 | 0.519 | 0.636 | 0.543 | 0.517 | 0.551 | 0.534 |
GBMIA[ | 0.649 | 0.600 | 0.597 | 0.506 | 0.501 | 0.655 | 0.612 | 0.608 | 0.517 | 0.517 |
本文模型(n=50) | 0.712 | 0.638 | 0.709 | 0.635 | 0.614 | 0.749 | 0.651 | 0.709 | 0.614 | 0.608 |
模型 | ILSVRC2012 | 花卉识别 | ||||||||
DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | |
黑盒1[ | 0.637 | 0.529 | 0.503 | 0.419 | 0.436 | 0.624 | 0.530 | 0.540 | 0.429 | 0.442 |
黑盒2[ | 0.631 | 0.566 | 0.511 | 0.423 | 0.478 | 0.622 | 0.546 | 0.543 | 0.419 | 0.504 |
黑盒3[ | 0.639 | 0.572 | 0.521 | 0.531 | 0.550 | 0.640 | 0.555 | 0.551 | 0.477 | 0.518 |
GAN-Leaks[ | 0.639 | 0.577 | 0.520 | 0.519 | 0.521 | 0.638 | 0.579 | 0.548 | 0.468 | 0.510 |
蒙特卡罗[ | 0.644 | 0.594 | 0.577 | 0.545 | 0.611 | 0.646 | 0.601 | 0.605 | 0.551 | 0.543 |
GBMIA[ | 0.657 | 0.608 | 0.600 | 0.547 | 0.630 | 0.655 | 0.603 | 0.594 | 0.548 | 0.540 |
本文模型(n=50) | 0.740 | 0.602 | 0.730 | 0.622 | 0.619 | 0.737 | 0.661 | 0.711 | 0.632 | 0.622 |
表2 在5个GAN下7类MIA的平均攻击成功率
Tab. 2 Average attack success rates of seven types of MIAs under five GANs
模型 | CIFAR10 | Fashion-MNIST | ||||||||
---|---|---|---|---|---|---|---|---|---|---|
DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | |
黑盒1[ | 0.576 | 0.514 | 0.492 | 0.493 | 0.466 | 0.628 | 0.517 | 0.500 | 0.436 | 0.451 |
黑盒2[ | 0.566 | 0.509 | 0.498 | 0.510 | 0.506 | 0.616 | 0.518 | 0.516 | 0.435 | 0.514 |
黑盒3[ | 0.561 | 0.530 | 0.479 | 0.517 | 0.502 | 0.633 | 0.545 | 0.507 | 0.481 | 0.509 |
GAN-Leaks[ | 0.618 | 0.530 | 0.501 | 0.531 | 0.492 | 0.621 | 0.536 | 0.512 | 0.530 | 0.505 |
蒙特卡罗[ | 0.667 | 0.544 | 0.505 | 0.522 | 0.519 | 0.636 | 0.543 | 0.517 | 0.551 | 0.534 |
GBMIA[ | 0.649 | 0.600 | 0.597 | 0.506 | 0.501 | 0.655 | 0.612 | 0.608 | 0.517 | 0.517 |
本文模型(n=50) | 0.712 | 0.638 | 0.709 | 0.635 | 0.614 | 0.749 | 0.651 | 0.709 | 0.614 | 0.608 |
模型 | ILSVRC2012 | 花卉识别 | ||||||||
DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | DCGAN | InfoGAN | WGAN | PIGAN | RS-DPGAN | |
黑盒1[ | 0.637 | 0.529 | 0.503 | 0.419 | 0.436 | 0.624 | 0.530 | 0.540 | 0.429 | 0.442 |
黑盒2[ | 0.631 | 0.566 | 0.511 | 0.423 | 0.478 | 0.622 | 0.546 | 0.543 | 0.419 | 0.504 |
黑盒3[ | 0.639 | 0.572 | 0.521 | 0.531 | 0.550 | 0.640 | 0.555 | 0.551 | 0.477 | 0.518 |
GAN-Leaks[ | 0.639 | 0.577 | 0.520 | 0.519 | 0.521 | 0.638 | 0.579 | 0.548 | 0.468 | 0.510 |
蒙特卡罗[ | 0.644 | 0.594 | 0.577 | 0.545 | 0.611 | 0.646 | 0.601 | 0.605 | 0.551 | 0.543 |
GBMIA[ | 0.657 | 0.608 | 0.600 | 0.547 | 0.630 | 0.655 | 0.603 | 0.594 | 0.548 | 0.540 |
本文模型(n=50) | 0.740 | 0.602 | 0.730 | 0.622 | 0.619 | 0.737 | 0.661 | 0.711 | 0.632 | 0.622 |
网络层 | ASR在50%以下 | ASR在50%~60% | ASR在60%以上 |
---|---|---|---|
FC1 | 0.34 | 0.71 | 0.83 |
FC2 | 0.13 | 0.85 | 0.83 |
FC3 | 0.09 | 0.78 | 0.87 |
Deconv1 | 0.41 | 0.67 | 0.70 |
Deconv2 | 0.33 | 0.65 | 0.76 |
BN1 | 0.51 | 0.52 | 0.49 |
BN2 | 0.57 | 0.52 | 0.55 |
所有层 | 0.30 | 0.69 | 0.95 |
表3 WGAN上不同网络层与攻击成功率的斯皮尔曼相关性
Tab. 3 Spearman’s correlation of different network layers and attack success rates on WGAN
网络层 | ASR在50%以下 | ASR在50%~60% | ASR在60%以上 |
---|---|---|---|
FC1 | 0.34 | 0.71 | 0.83 |
FC2 | 0.13 | 0.85 | 0.83 |
FC3 | 0.09 | 0.78 | 0.87 |
Deconv1 | 0.41 | 0.67 | 0.70 |
Deconv2 | 0.33 | 0.65 | 0.76 |
BN1 | 0.51 | 0.52 | 0.49 |
BN2 | 0.57 | 0.52 | 0.55 |
所有层 | 0.30 | 0.69 | 0.95 |
网络层 | ASR在50%以下 | ASR在50%~60% | ASR在60%以上 |
---|---|---|---|
Deconv1 | 0.46 | 0.85 | 0.74 |
Deconv2 | 0.37 | 0.84 | 0.89 |
Deconv3 | 0.27 | 0.81 | 0.86 |
BN1 | 0.47 | 0.56 | 0.51 |
BN2 | 0.50 | 0.52 | 0.58 |
所有层 | -0.11 | 0.71 | 0.92 |
表4 DCGAN上不同网络层与攻击成功率的斯皮尔曼相关性
Tab. 4 Spearman’s correlation of different network layers and attack success rates on DCGAN
网络层 | ASR在50%以下 | ASR在50%~60% | ASR在60%以上 |
---|---|---|---|
Deconv1 | 0.46 | 0.85 | 0.74 |
Deconv2 | 0.37 | 0.84 | 0.89 |
Deconv3 | 0.27 | 0.81 | 0.86 |
BN1 | 0.47 | 0.56 | 0.51 |
BN2 | 0.50 | 0.52 | 0.58 |
所有层 | -0.11 | 0.71 | 0.92 |
[1] | 李乐阳,佟国香,赵迎志,等. 基于生成对抗网络的文本生成图像研究综述[J]. 电子科技, 2023, 36(10): 39-55. |
LI Y Y, TONG G X, ZHAO Y Z, et al. A survey of text-to-image synthesis based on generative adversarial network[J]. Electronic Science and Technology, 2023, 36(10): 39-55. | |
[2] | 王崇宇,毛琪,金立标. 基于生成对抗网络的图像视频编码综述[J]. 中国传媒大学学报(自然科学版), 2022, 29(6): 19-28. |
WANG C Y, MAO Q, JIN L B. Review on image and video coding via generative adversarial networks[J]. Journal of Communication University of China (Science and Technology), 2022, 29(6): 19-28. | |
[3] | PRADHYUMNA P, MOHANA. A survey of modern deep learning based Generative Adversarial Networks (GANs)[C]// Proceedings of the 6th International Conference on Computing Methodologies and Communication. Piscataway: IEEE, 2022: 1146-1152. |
[4] | HU H, SALCIC Z, SUN L, et al. Membership inference attacks on machine learning: a survey[J]. ACM Computing Surveys, 2022, 54(11s): No.235. |
[5] | 牛俊,马骁骥,陈颖,等. 机器学习中成员推理攻击和防御研究综述[J].信息安全学报, 2022, 7(6): 1-30. |
NIU J, MA X J, CHEN Y, et al. A survey on membership inference attacks and defenses in machine learning[J]. Journal of Cyber Security, 2022, 7(6): 1-30. | |
[6] | CHEN D, YU N, ZHANG Y, et al. GAN-Leaks: a taxonomy of membership inference attacks against GANs[C]// Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2020:343-362. |
[7] | HILPRECHT B, HERTERICH M, BERNAU D. Monte Carlo and reconstruction membership inference attacks against generative models[J]. Proceedings on Privacy Enhancing Technologies, 2019(4):232-249. |
[8] | WANG X, WANG N, WU L, et al. GBMIA: gradient-based membership inference attack in federated learning[C]// Proceedings of the 2023 IEEE International Conference on Communications. Piscataway: IEEE, 2023: 5066-5071. |
[9] | 彭长根,高婷,刘惠篮,等. 面向机器学习模型的基于PCA的成员推理攻击[J]. 通信学报, 2022, 43(1): 149-160. |
PENG C G, GAO T, LIU H L, et al. PCA-based membership inference attack for machine learning models[J]. Journal on Communications, 2022, 43(1): 149-160. | |
[10] | ZHANG M, YU N, WEN R, et al. Generated distributions are all you need for membership inference attacks against generative models[C]// Proceedings of the 2024 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway: IEEE, 2024:4827-4837. |
[11] | ZHANG Y, ZHOU H, WANG P, et al. Black-box based limited query membership inference attack[J]. IEEE Access, 2022, 10: 55459-55468. |
[12] | BLEI D M, KUCUKELBIR A, McAULIFFE J D. Variational inference: a review for statisticians[J]. Journal of the American Statistical Association, 2017, 13(112): No.859877. |
[13] | RADFORD A, METZ L, CHINTALA S. Unsupervised representation learning with deep convolutional generative adversarial networks[EB/OL]. [2024-09-12].. |
[14] | SHOKRI R, STRONATI M, SONG C, et al. Membership inference attacks against machine learning models[C]// Proceedings of the 2017 IEEE Symposium on Security and Privacy. Piscataway: IEEE, 2017: 3-18. |
[15] | WU Y, BURDA Y, SALAKHUTDINOV R, et al. On the quantitative analysis of decoder-based generative models[EB/OL]. [2024-07-08].. |
[16] | YEOM S, GIACOMELLI I, FREDRIKSON M, et al. Privacy risk in machine learning: analyzing the connection to overfitting[C]// Proceedings of the IEEE 31st Computer Security Foundations Symposium. Piscataway: IEEE, 2018: 268-282. |
[17] | NASR M, SHOKRI R, HOUMANSADR A. Comprehensive privacy analysis of deep learning: passive and active white-box inference attacks against centralized and federated learning[C]// Proceedings of the 2019 IEEE Symposium on Security and Privacy. Piscataway: IEEE, 2019: 739-753. |
[18] | HU H, PANG J. Membership inference attacks against GANs by leveraging over-representation regions[C]// Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. New York: ACM, 2021: 2387-2389. |
[19] | ZHENG J, CAO Y, WANG H. Resisting membership inference attacks through knowledge distillation[J]. Neurocomputing, 2021, 452: 114-126. |
[20] | EFRON B. Bayes’ theorem in the 21st century[J]. Science, 2013, 240(6137): 1177-1178. |
[21] | STIGLER S M. Thomas Bayes’s Bayesian inference[J]. Journal of the Royal Statistical Society. Series A (General), 1982, 145(2):250-258. |
[22] | JORDAN M I, GHAHRAMANI Z, JAAKKOLA T S, et al. An introduction to variational methods for graphical models[J]. Machine Learning, 1999, 37(2): 183-233. |
[23] | ISHIGURO K, SATO I, UEDA N. Averaged collapsed variational Bayes inference[J]. Journal of Machine Learning Research, 2017, 18: 1-29. |
[24] | BISHOP C M. Pattern recognition and machine learning[M]. New York: Springer, 2006. |
[25] | XIAO H, RASUL K, VOLLGRAF R. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms[EB/OL]. [2024-09-12].. |
[26] | KRIZHEVSKY A. Learning multiple layers of features from tiny images[R/OL]. [2024-09-12].. |
[27] | RUSSAKOVSKY O, DENG J, SU H, et al. ImageNet large scale visual recognition challenge[J]. International Journal of Computer Vision, 2015, 115(3): 211-252. |
[28] | NILSBACK M E, ZISSERMAN A. Automated flower classification over a large number of classes[C]// Proceedings of the 6th Indian Conference on Computer Vision, Graphics and Image Processing. Piscataway: IEEE, 2008: 722-729. |
[29] | CHEN X, DUAN Y, HOUTHOOFT R, et al. InfoGAN: interpretable representation learning by information maximizing generative adversarial nets[C]// Proceedings of the 30th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2016: 2180-2188. |
[30] | MARTIN A, SOUMITH C, LEON B. Wasserstein generative adversarial networks[C]// Proceedings of the 34th International Conference on Machine Learning. New York: JMLR.org, 2017: 214-223. |
[31] | HASSANZADEH P H, TILLMAN R E. Generative models with information-theoretic protection against membership inference attacks[EB/OL]. [2024-10-01].. |
[32] | HUANG Y, CAO L. Privacy-preserving remote sensing image generation and classification with differentially private GANs[J]. IEEE Sensors Journal, 2023, 23(18): 20805-20816. |
[1] | 梁永濠, 李金龙. 用于神经布尔可满足性问题求解器的新型消息传递网络[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2934-2940. |
[2] | 葛丽娜, 王明禹, 田蕾. 联邦学习的高效性研究综述[J]. 《计算机应用》唯一官方网站, 2025, 45(8): 2387-2398. |
[3] | 郭书剑, 余节约, 尹学松. 图正则化弹性网子空间聚类[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1464-1471. |
[4] | 朱俊屹, 常雷雷, 徐晓滨, 郝智勇, 于海跃, 姜江. 基于最小先验知识的自监督学习方法[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1035-1041. |
[5] | 洪梓榕, 包广清. 基于集成学习的雷达自动目标识别综述[J]. 《计算机应用》唯一官方网站, 2025, 45(2): 371-382. |
[6] | 姚梓豪, 栗远明, 马自强, 李扬, 魏良根. 基于机器学习的多目标缓存侧信道攻击检测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1862-1871. |
[7] | 陈学斌, 任志强, 张宏扬. 联邦学习中的安全威胁与防御措施综述[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1663-1672. |
[8] | 郑毅, 廖存燚, 张天倩, 王骥, 刘守印. 面向城区的基于图去噪的小区级RSRP估计方法[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 855-862. |
[9] | 佘维, 李阳, 钟李红, 孔德锋, 田钊. 基于改进实数编码遗传算法的神经网络超参数优化[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 671-676. |
[10] | 任帅, 纪元法, 孙希延, 韦照川, 林子安. 基于改进灰狼优化与支持向量回归的滑坡位移预测[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 972-982. |
[11] | 余孙婕, 曾辉, 熊诗雨, 史红周. 基于生成式对抗网络的联邦学习激励机制[J]. 《计算机应用》唯一官方网站, 2024, 44(2): 344-352. |
[12] | 李博, 黄建强, 黄东强, 王晓英. 基于异构平台的稀疏矩阵向量乘自适应计算优化[J]. 《计算机应用》唯一官方网站, 2024, 44(12): 3867-3875. |
[13] | 陈学斌, 屈昌盛. 面向联邦学习的后门攻击与防御综述[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3459-3469. |
[14] | 孙仁科, 皇甫志宇, 陈虎, 李仲年, 许新征. 神经架构搜索综述[J]. 《计算机应用》唯一官方网站, 2024, 44(10): 2983-2994. |
[15] | 柴汶泽, 范菁, 孙书魁, 梁一鸣, 刘竟锋. 深度度量学习综述[J]. 《计算机应用》唯一官方网站, 2024, 44(10): 2995-3010. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||