Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks

Membership inference (MI) attacks threaten user privacy through determining if a given data example has been used to train a target model. Existing MI defenses protect the membership privacy through preemptive exclusion of members techniques and knowledge distillation. Unfortunately, using either of...

Full description

Saved in:
Bibliographic Details
Main Authors: Jun Niu, Peng Liu, Chunhui Huang, Yangming Zhang, Moxuan Zeng, Kuo Shen, Yangzhong Wang, Suyu An, Yulong Shen, Xiaohong Jiang, Jianfeng Ma, He Wang, Gaofei Wu, Anmin Fu, Chunjie Cao, Xiaoyan Zhu, Yuqing Zhang
Format: Article
Language:English
Published: KeAi Communications Co., Ltd. 2025-01-01
Series:Journal of Information and Intelligence
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2949715924000556
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850054873125486592
author Jun Niu
Peng Liu
Chunhui Huang
Yangming Zhang
Moxuan Zeng
Kuo Shen
Yangzhong Wang
Suyu An
Yulong Shen
Xiaohong Jiang
Jianfeng Ma
He Wang
Gaofei Wu
Anmin Fu
Chunjie Cao
Xiaoyan Zhu
Yuqing Zhang
author_facet Jun Niu
Peng Liu
Chunhui Huang
Yangming Zhang
Moxuan Zeng
Kuo Shen
Yangzhong Wang
Suyu An
Yulong Shen
Xiaohong Jiang
Jianfeng Ma
He Wang
Gaofei Wu
Anmin Fu
Chunjie Cao
Xiaoyan Zhu
Yuqing Zhang
author_sort Jun Niu
collection DOAJ
description Membership inference (MI) attacks threaten user privacy through determining if a given data example has been used to train a target model. Existing MI defenses protect the membership privacy through preemptive exclusion of members techniques and knowledge distillation. Unfortunately, using either of these two defenses alone, the defense effect can still offers an unsatisfactory trade-off between membership privacy and utility.Given that the defense method that directly combines these two defenses is still very limited (e.g., the test accuracy of the target model is decreased by about 40% (in our experiments)), in this work, we propose a dual defense (DD) method that includes the preemptive exclusion of high-risk member samples module and the knowledge distillation module, which thwarts the access of the resulting models to the private training data twice to mitigate MI attacks. Our defense method can be divided into two steps: the preemptive exclusion of high-risk member samples (Step 1) and the knowledge distillation to obtain the protected student model (Step 2). We propose three types of exclusions: existing MI attacks-based exclusions, sample distances of members and nonmembers-based exclusions, and mutual information value-based exclusions, to preemptively exclude the high-risk member samples. During the knowledge distillation phase, we add ground-truth labeled data to the reference dataset to decrease the protected student model's dependency on soft labels, aiming to maintain or improve its test accuracy. Extensive evaluation shows that DD significantly outperforms state-of-the-art defenses and offers a better privacy-utility trade-off. For example, DD achieves ∼100% test accuracy improvement over the distillation for membership privacy (DMP) defense for ResNet50 trained on CIFAR100. DD simultaneously achieves the reductions in the attack effectiveness (e.g., the TPR@0.01%FPR of enhanced MI attacks decreased by 2.10% on the ImageNet dataset, the membership advantage (MA) of risk score-based attacks decreased by 56.30%) and improvements of the target models' test accuracies (e.g., by 42.80% on CIFAR100).
format Article
id doaj-art-74bdb418416448f1bdc8d8df548175b9
institution DOAJ
issn 2949-7159
language English
publishDate 2025-01-01
publisher KeAi Communications Co., Ltd.
record_format Article
series Journal of Information and Intelligence
spelling doaj-art-74bdb418416448f1bdc8d8df548175b92025-08-20T02:52:08ZengKeAi Communications Co., Ltd.Journal of Information and Intelligence2949-71592025-01-0131689010.1016/j.jiixd.2024.06.002Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacksJun Niu0Peng Liu1Chunhui Huang2Yangming Zhang3Moxuan Zeng4Kuo Shen5Yangzhong Wang6Suyu An7Yulong Shen8Xiaohong Jiang9Jianfeng Ma10He Wang11Gaofei Wu12Anmin Fu13Chunjie Cao14Xiaoyan Zhu15Yuqing Zhang16School of Computer Science and Technology, Xidian University, Xi'an 710071, ChinaInformation Sciences and Technology, Pennsylvania State University, PA 16802, USASchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaSchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaSchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaSchool of Computer Science and Technology, Xidian University, Xi'an 710071, China; Information Sciences and Technology, Pennsylvania State University, PA 16802, USA; School of Cyberspace Security, Hainan University, Haikou 570228, China; School of Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China; School of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, Japan; School of Cyber Engineering, Xidian University, Xi'an 710071, China; School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China; School of Telecommunications Engineering, Xidian University, Xi'an 710071, China; School of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, ChinaSchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaSchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaSchool of Computer Science and Technology, Xidian University, Xi'an 710071, ChinaSchool of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, JapanSchool of Computer Science and Technology, Xidian University, Xi'an 710071, China; Information Sciences and Technology, Pennsylvania State University, PA 16802, USA; School of Cyberspace Security, Hainan University, Haikou 570228, China; School of Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China; School of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, Japan; School of Cyber Engineering, Xidian University, Xi'an 710071, China; School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China; School of Telecommunications Engineering, Xidian University, Xi'an 710071, China; School of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, ChinaSchool of Computer Science and Technology, Xidian University, Xi'an 710071, China; Information Sciences and Technology, Pennsylvania State University, PA 16802, USA; School of Cyberspace Security, Hainan University, Haikou 570228, China; School of Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China; School of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, Japan; School of Cyber Engineering, Xidian University, Xi'an 710071, China; School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China; School of Telecommunications Engineering, Xidian University, Xi'an 710071, China; School of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, ChinaSchool of Computer Science and Technology, Xidian University, Xi'an 710071, China; Information Sciences and Technology, Pennsylvania State University, PA 16802, USA; School of Cyberspace Security, Hainan University, Haikou 570228, China; School of Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China; School of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, Japan; School of Cyber Engineering, Xidian University, Xi'an 710071, China; School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China; School of Telecommunications Engineering, Xidian University, Xi'an 710071, China; School of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, ChinaSchool of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, ChinaSchool of Cyberspace Security, Hainan University, Haikou 570228, ChinaCorresponding authors.; School of Computer Science and Technology, Xidian University, Xi'an 710071, China; Information Sciences and Technology, Pennsylvania State University, PA 16802, USA; School of Cyberspace Security, Hainan University, Haikou 570228, China; School of Hangzhou Institute of Technology, Xidian University, Hangzhou 311231, China; School of Systems Information Science, Future University of Hakodate, Hakodate 041-8655, Japan; School of Cyber Engineering, Xidian University, Xi'an 710071, China; School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China; School of Telecommunications Engineering, Xidian University, Xi'an 710071, China; School of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, ChinaSchool of Computer & Control Engineering, University of Chinese Academy of Sciences, Beijing 101408, China; Corresponding authors.Membership inference (MI) attacks threaten user privacy through determining if a given data example has been used to train a target model. Existing MI defenses protect the membership privacy through preemptive exclusion of members techniques and knowledge distillation. Unfortunately, using either of these two defenses alone, the defense effect can still offers an unsatisfactory trade-off between membership privacy and utility.Given that the defense method that directly combines these two defenses is still very limited (e.g., the test accuracy of the target model is decreased by about 40% (in our experiments)), in this work, we propose a dual defense (DD) method that includes the preemptive exclusion of high-risk member samples module and the knowledge distillation module, which thwarts the access of the resulting models to the private training data twice to mitigate MI attacks. Our defense method can be divided into two steps: the preemptive exclusion of high-risk member samples (Step 1) and the knowledge distillation to obtain the protected student model (Step 2). We propose three types of exclusions: existing MI attacks-based exclusions, sample distances of members and nonmembers-based exclusions, and mutual information value-based exclusions, to preemptively exclude the high-risk member samples. During the knowledge distillation phase, we add ground-truth labeled data to the reference dataset to decrease the protected student model's dependency on soft labels, aiming to maintain or improve its test accuracy. Extensive evaluation shows that DD significantly outperforms state-of-the-art defenses and offers a better privacy-utility trade-off. For example, DD achieves ∼100% test accuracy improvement over the distillation for membership privacy (DMP) defense for ResNet50 trained on CIFAR100. DD simultaneously achieves the reductions in the attack effectiveness (e.g., the TPR@0.01%FPR of enhanced MI attacks decreased by 2.10% on the ImageNet dataset, the membership advantage (MA) of risk score-based attacks decreased by 56.30%) and improvements of the target models' test accuracies (e.g., by 42.80% on CIFAR100).http://www.sciencedirect.com/science/article/pii/S2949715924000556Machine learningMembership inference defensesPreemptive exclusionKnowledge distillation
spellingShingle Jun Niu
Peng Liu
Chunhui Huang
Yangming Zhang
Moxuan Zeng
Kuo Shen
Yangzhong Wang
Suyu An
Yulong Shen
Xiaohong Jiang
Jianfeng Ma
He Wang
Gaofei Wu
Anmin Fu
Chunjie Cao
Xiaoyan Zhu
Yuqing Zhang
Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
Journal of Information and Intelligence
Machine learning
Membership inference defenses
Preemptive exclusion
Knowledge distillation
title Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
title_full Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
title_fullStr Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
title_full_unstemmed Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
title_short Dual defense: Combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
title_sort dual defense combining preemptive exclusion of members and knowledge distillation to mitigate membership inference attacks
topic Machine learning
Membership inference defenses
Preemptive exclusion
Knowledge distillation
url http://www.sciencedirect.com/science/article/pii/S2949715924000556
work_keys_str_mv AT junniu dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT pengliu dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT chunhuihuang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT yangmingzhang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT moxuanzeng dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT kuoshen dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT yangzhongwang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT suyuan dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT yulongshen dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT xiaohongjiang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT jianfengma dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT hewang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT gaofeiwu dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT anminfu dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT chunjiecao dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT xiaoyanzhu dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks
AT yuqingzhang dualdefensecombiningpreemptiveexclusionofmembersandknowledgedistillationtomitigatemembershipinferenceattacks