Article(id=1329420702152766421, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1329420635937288321, articleNumber=null, orderNo=null, doi=10.7544/issn1000-1239.202660484, pmid=null, cstr=null, oa=null, hot=null, price=null, onlineType=0, articleFormat=0, articleType=null, articleTypeStr=null, receivedDate=1780675200000, receivedDateStr=2026-06-06, revisedDate=1782921600000, revisedDateStr=2026-07-02, acceptedDate=null, acceptedDateStr=null, onlineDate=1794888194834, onlineDateStr=2026-11-17, pubDate=1790784000000, pubDateStr=2026-10-01, doiRegisterDate=null, doiRegisterDateStr=null, onlineIssueDate=1794888194834, onlineIssueDateStr=2026-11-17, onlineJustAcceptDate=null, onlineJustAcceptDateStr=null, onlineFirstDate=null, onlineFirstDateStr=null, sourceXml=null, magXml=null, createTime=1794888194834, creator=13701087609, updateTime=1794888194834, updator=13701087609, issue=Issue{id=1329420635937288321, tenantId=1146029695717560320, journalId=1324759914146529478, year='2026', volume='63', issue='10', pageStart='2479', pageEnd='2766', issueExtLink='null', onlineDate='null', pubDate='1790784000000', pubDateStr='2026-10-01', beforeIssueId=null, nextIssueId=null, price=null, status=1, issueComplete=1, articleOrder=1, issueType=-1, specialIssue=null, createTime=1794888179048, creator='13701087609', updateTime=1794888179048, updator='13701087609', preIssue=null, nextIssue=null, articleTotal=null, ext=null, issueFiles=null, downloadFileDto=null}, startPage=2576, endPage=2595, ext={EN=ArticleExt(id=1329420702358287318, articleId=1329420702152766421, tenantId=1146029695717560320, journalId=1324759914146529478, language=EN, title=Byzantine-Robust Federated Learning for Network Intrusion Detection via Gradient Fingerprint-Driven Decision, columnId=1329420636595794050, journalTitle=Journal of Computer Research and Development, columnName=Special Issue on Security Protection Technology of Massive Heterogeneous Terminals in Dynamic Heterogeneous Networks, runingTitle=null, highlight=null, articleAbstract=

In dynamic heterogeneous networks, massive heterogeneous terminals result in dispersed and diverse security data. While federated learning enables collaborative intrusion detection across terminals without compromising data privacy, its open architecture is inherently vulnerable to Byzantine attacks, where malicious clients poison the global model by uploading crafted gradients. Existing Byzantine-robust methods are typically effective only against specific attack types and struggle to cope with the coexistence of multiple attacks in real-world environments. Moreover, they suffer from high false-negative rates under sparse real-traffic conditions, and the computational overhead of stacking multiple defenses is prohibitive for resource-constrained terminals. To address these challenges, we propose GF-FedIDS, a Byzantine-robust federated learning intrusion detection framework based on gradient fingerprint decision-making. By analyzing the gradient statistics of typical Byzantine attacks, we establish a deterministic mapping from attack patterns to optimal defenses and reveal their causal relationship. A gradient fingerprint decision tree is designed to adaptively select a single optimal defense strategy, achieving full-coverage robust aggregation at unit computational cost. Furthermore, a class-aware instance weighting mechanism is introduced to mitigate missed detections under data imbalance. Extensive experiments on three public datasets and eight attack types demonstrate that, in single-attack scenarios, the attack detection rate (ADR) of GF-FedIDS outperforms that of the best baseline across three datasets; in mixed-attack scenarios, ADR reaches 1.45–2.14 times; and under extreme imbalance with a 5% attack rate, ADR is 2.06–2.29 times that of the best baseline. Additionally, GF-FedIDS incurs only a 16% computational overhead compared with FedAvg, outperforming existing ensemble defense methods.

, authors=Zhiyuan Jiang1, Gang Yang2, *, Yongjun Wang1, Zheng Qin3, authorsList=Zhiyuan Jiang, Gang Yang, Yongjun Wang, Zheng Qin, authorCompany=null, correspAuthors=Gang Yang, authorNote=

born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

, correspAuthorsNote=null, copyrightStatement=Copyright © 2026 Journal of Computer Research and Development. All rights reserved., copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=null, magXml=null, pdfUrl=null, pdf=null, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=null, mapNumber=null, fund=null), CN=ArticleExt(id=1329420704283472873, articleId=1329420702152766421, tenantId=1146029695717560320, journalId=1324759914146529478, language=CN, title=基于梯度指纹决策的拜占庭鲁棒联邦学习网络入侵检测方法, columnId=1329420636847452292, journalTitle=计算机研究与发展, columnName=动态异构网络中海量异质终端安全防护技术专题, runingTitle=null, highlight=null, articleAbstract=

动态异构网络中,海量异质终端导致安全数据分散异构,联邦学习虽能保护数据隐私实现跨终端的协同入侵检测,但其开放架构易受拜占庭攻击,即恶意客户端通过上传恶意梯度破坏全局模型。现有拜占庭鲁棒方法通常仅在特定攻击类型上有效,难以覆盖真实场景中多类型攻击共存的态势,且在攻击样本稀疏的真实流量环境下易产生大量漏报,多防御堆叠的计算开销亦难以适配资源受限终端。对此,提出基于梯度指纹决策的拜占庭鲁棒联邦学习入侵检测框架GF-FedIDS(gradient fingerprint-driven federated intrusion detection system),通过对典型拜占庭攻击梯度统计特性的分析,建立攻击模式到最优防御的确定性映射,揭示其因果关联;设计梯度指纹决策树,实现自适应选择单一最优防御策略,以单位计算开销达成全覆盖鲁棒聚合;引入类别感知实例加权机制,缓解数据不平衡下的漏报问题。在3个公开数据集、8种攻击类型条件下的广泛实验表明,在单攻击场景下,GF-FedIDS的攻击检测率(ADR)在跨数据集场景下普遍略优于最优基线;在混合攻击场景下,ADR为最优基线的1.45~2.14倍;在攻击率为5%的极端不平衡条件下,ADR为最优基线的2.06~2.29倍。此外,GF-FedIDS与FedAvg相比仅增加16%的计算开销,优于现有集成防御方法。

, authors=姜植元1, 杨刚2, *, 王勇军1, 秦拯3, authorsList=姜植元, 杨刚, 王勇军, 秦拯, authorCompany=null, correspAuthors=杨刚, authorNote=

作者贡献声明

姜植元提出了研究方法与解决思路,以及方法论的整体构想;杨刚负责具体设计实现并撰写论文;王勇军、秦拯为论文提出指导意见。

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

, correspAuthorsNote=
杨刚()
, copyrightStatement=版权所有 © 2026 计算机研究与发展编辑部, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=+v1qw/ZxSTP27VAyt+hgNg==, magXml=X8LE/uWEpHVyAbFT5ZExYA==, pdfUrl=null, pdf=40gfd0LGshoUFH3KCM0Bbw==, pdfFileSize=2257335, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=0xJs81KT7DaZOBcIG91h5w==, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=xMWB1xuG2kCYfZivSlZrTg==, mapNumber=null, fund=null)}, authors=[Author(id=1329420704824538100, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=jzy@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420704904229878, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420704824538100, language=EN, stringName=Zhiyuan Jiang, firstName=Zhiyuan, middleName=null, lastName=Jiang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

"}, bioImg=null, bioContent=

born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420704967144439, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420704824538100, language=CN, stringName=姜植元, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1国防科技大学计算机学院 长沙 410073, bio={"img":"ArsWPHysmqF0+Pm30DbQhA==","content":"

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

"}, bioImg=ArsWPHysmqF0+Pm30DbQhA==, bioContent=

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704539325418, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=1, ext=[AuthorCompanyExt(id=1329420704572879851, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1329420704606434284, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1国防科技大学计算机学院 长沙 410073)])]), Author(id=1329420705025864697, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=yanggang11@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329420705113945083, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705025864697, language=EN, stringName=Gang Yang, firstName=Gang, middleName=null, lastName=Yang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, *, address=2College of Network and Information, Information Support Force Engineering University, Wuhan 430030, bio={"content":"

born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

"}, bioImg=null, bioContent=

born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705189442556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705025864697, language=CN, stringName=杨刚, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, *, address=2信息支援部队工程大学网络信息学院 武汉 430030, bio={"img":"yHBJbn6/ehSweB35Ac677Q==","content":"

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

"}, bioImg=yHBJbn6/ehSweB35Ac677Q==, bioContent=

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704677737453, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=2, ext=[AuthorCompanyExt(id=1329420704681931758, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2College of Network and Information, Information Support Force Engineering University, Wuhan 430030), AuthorCompanyExt(id=1329420704690320367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2信息支援部队工程大学网络信息学院 武汉 430030)])]), Author(id=1329420705264940030, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420705373991936, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705264940030, language=EN, stringName=Yongjun Wang, firstName=Yongjun, middleName=null, lastName=Wang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

"}, bioImg=null, bioContent=

born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705445294080, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705264940030, language=CN, stringName=王勇军, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1国防科技大学计算机学院 长沙 410073, bio={"img":"HgQj/LSlTautaN/aYy12kg==","content":"

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

"}, bioImg=HgQj/LSlTautaN/aYy12kg==, bioContent=

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704539325418, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=1, ext=[AuthorCompanyExt(id=1329420704572879851, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1329420704606434284, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1国防科技大学计算机学院 长沙 410073)])]), Author(id=1329420705508208642, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420705579511812, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705508208642, language=EN, stringName=Zheng Qin, firstName=Zheng, middleName=null, lastName=Qin, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3College of Cyber Science and Technology, Hunan University, Changsha 410082, bio={"content":"

born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

"}, bioImg=null, bioContent=

born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705663397893, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705508208642, language=CN, stringName=秦拯, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3湖南大学网络空间安全学院 长沙 410082, bio={"img":"Y68ndMp8waECsNVk3tfqpQ==","content":"

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

"}, bioImg=Y68ndMp8waECsNVk3tfqpQ==, bioContent=

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704749040624, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=3, ext=[AuthorCompanyExt(id=1329420704757429233, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Cyber Science and Technology, Hunan University, Changsha 410082), AuthorCompanyExt(id=1329420704765817842, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3湖南大学网络空间安全学院 长沙 410082)])])], keywords=[Keyword(id=1329420705789227014, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, orderNo=1, keyword=federated learning), Keyword(id=1329420705881501703, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, orderNo=2, keyword=Byzantine attack), Keyword(id=1329420705998942216, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, orderNo=3, keyword=intrusion detection), Keyword(id=1329420706078633993, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, orderNo=4, keyword=gradient fingerprint), Keyword(id=1329420706154131466, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, orderNo=5, keyword=Non-IID data), Keyword(id=1329420706233823243, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, orderNo=1, keyword=联邦学习), Keyword(id=1329420706317709324, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, orderNo=2, keyword=拜占庭攻击), Keyword(id=1329420706422566925, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, orderNo=3, keyword=入侵检测), Keyword(id=1329420706548396046, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, orderNo=4, keyword=梯度指纹), Keyword(id=1329420706602921999, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, orderNo=5, keyword=非独立同分布数据)], refs=[Reference(id=1329420708410667046, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=0, authorNames=null, journalName=null, refType=null, unstructuredReference=Xu Ke, Yi Qiu, Kun Yang. A tutorial on the Internet of things: From a heterogeneous network integration perspective[J]. IEEE Network, 2016, 30(2): 102−108, articleTitle=null, refAbstract=null), Reference(id=1329420708578439207, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=2, rfOrder=1, authorNames=null, journalName=null, refType=null, unstructuredReference=Brendan M, Eider M, Daniel R, et al. Communication-efficient learning of deep networks from decentralized data[C]//Proc of the 20th Int Conf on Artificial Intelligence and Statistics. Fort Lauderdale, New York: PMLR, 2017: 1273−1282, articleTitle=null, refAbstract=null), Reference(id=1329420708645548072, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=3, rfOrder=2, authorNames=null, journalName=null, refType=null, unstructuredReference=Amine F M, Othmane F, Leandros M, et al. Federated deep learning for cyber security in the Internet of things: Concepts, applications, and experimental analysis[J]. IEEE Access, 2021, 9: 138509−138542, articleTitle=null, refAbstract=null), Reference(id=1329420708725239849, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=4, rfOrder=3, authorNames=null, journalName=null, refType=null, unstructuredReference=Nguyen D C, Ding Ming, Pathirana P N, et al. Federated learning for Internet of things: A comprehensive survey[J]. IEEE Communications Surveys & Tutorials, 2021, 23(3): 1622−1658, articleTitle=null, refAbstract=null), Reference(id=1329420708876234794, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=5, rfOrder=4, authorNames=null, journalName=null, refType=null, unstructuredReference=de Oliveira J A, Gonçalves V P, Meneguette R I, et al. F-NIDS—A network intrusion detection system based on federated learning[J]. Computer Networks, 2023, 236: 110010, articleTitle=null, refAbstract=null), Reference(id=1329420708976898091, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=6, rfOrder=5, authorNames=null, journalName=null, refType=null, unstructuredReference=刘飚, 张方佼, 王文鑫, 等. 基于矩阵映射的拜占庭鲁棒联邦学习算法[J]. 计算机研究与发展, 2021, 58(11): 2416−2429, articleTitle=null, refAbstract=null), Reference(id=1329420709048201260, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=6, rfOrder=6, authorNames=null, journalName=null, refType=null, unstructuredReference=Liu Biao, Zhang Fangjiao, Wang Wenxin, et al. A Byzantine-robust federated learning algorithm based on matrix mapping[J]. Journal of Computer Research and Development, 2021, 58(11): 2416−2429 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420709153058861, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=7, rfOrder=7, authorNames=null, journalName=null, refType=null, unstructuredReference=Xie C, Koyejo O, Gupta I. Fall of empires: Breaking Byzantine-tolerant SGD by inner product manipulation[C]//Proc of the 36th Conf on Uncertainty in Artificial Intelligence (UAI). Toronto, Canada: PMLR, 2020: 261−270, articleTitle=null, refAbstract=null), Reference(id=1329420709249527854, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=8, rfOrder=8, authorNames=null, journalName=null, refType=null, unstructuredReference=Blanchard P, El Mhamdi E M, Guerraoui R, et al. Machine learning with adversaries: Byzantine tolerant gradient descent[C]//Proc of the 31st Conf on Neural Information Processing Systems (NIPS). New York: Curran Associates, 2017: 119−129, articleTitle=null, refAbstract=null), Reference(id=1329420709329219631, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=9, rfOrder=9, authorNames=null, journalName=null, refType=null, unstructuredReference=Yin D, Chen Y, Kannan R, et al. Byzantine-robust distributed learning: Towards optimal statistical rates[C]//Proc of the 35th Int Conf on Machine Learning (ICML). Stockholm, Sweden: PMLR, 2018: 5650−5659, articleTitle=null, refAbstract=null), Reference(id=1329420709404717104, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=10, rfOrder=10, authorNames=null, journalName=null, refType=null, unstructuredReference=Muñoz-González L, Co K T, Lupu E C. Byzantine-robust federated machine learning through adaptive model averaging[DB/OL]. arXiv preprint, arXiv: 1909.05125, 2019, articleTitle=null, refAbstract=null), Reference(id=1329420709480214578, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=11, rfOrder=11, authorNames=null, journalName=null, refType=null, unstructuredReference=Li T, Sahu A K, Zaheer M, et al. Federated optimization in heterogeneous networks[C]//Proc of the 2nd Conf on Machine Learning and Systems (MLSys). Austin, TX: mlsys. org, 2020: 429−450, articleTitle=null, refAbstract=null), Reference(id=1329420709551517747, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=12, authorNames=null, journalName=null, refType=null, unstructuredReference=莫慧凌, 郑海峰, 高敏, 等. 基于联邦学习的多源异构数据融合算法[J]. 计算机研究与发展, 2022, 59(2): 478−487, articleTitle=null, refAbstract=null), Reference(id=1329420709622820916, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=13, authorNames=null, journalName=null, refType=null, unstructuredReference=Mo Huiling, Zheng Haifeng, Gao Min, et al. Multi-source heterogeneous data fusion based on federated learning[J]. Journal of Computer Research and Development, 2022, 59(2): 478−487 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420709715095605, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=13, rfOrder=14, authorNames=null, journalName=null, refType=null, unstructuredReference=Guerraoui R, Rouault S. The hidden vulnerability of distributed learning in Byzantium[C]//Proc of the 35th Int Conf on Machine Learning (ICML). Stockholm, Sweden: PMLR, 2018: 3521−3530, articleTitle=null, refAbstract=null), Reference(id=1329420709798981687, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=14, rfOrder=15, authorNames=null, journalName=null, refType=null, unstructuredReference=Dritsas E, Trigka M. Federated learning for IoT: A survey of techniques, challenges, and applications[J]. Journal of Sensor and Actuator Networks, 2025, 14(1): 9, articleTitle=null, refAbstract=null), Reference(id=1329420709878673464, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=15, rfOrder=16, authorNames=null, journalName=null, refType=null, unstructuredReference=Xiang Hanlong, Zhang Xuyun, Xu Xiaolong, et al. Federated learning-based anomaly detection with isolation forest in the IoT-edge continuum[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2026, 22(1): 1−19, articleTitle=null, refAbstract=null), Reference(id=1329420709954170937, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=16, rfOrder=17, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Zikai, Suman R, Xu Jiahao, et al. Federated learning for smart grid: A survey on applications and potential vulnerabilities[J]. ACM Transactions on Cyber-Physical Systems, 2026, 10(1): 1−26, articleTitle=null, refAbstract=null), Reference(id=1329420710025474106, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=17, rfOrder=18, authorNames=null, journalName=null, refType=null, unstructuredReference=Alhasawi Y, Alghamdi S. Federated learning for decentralized DDoS attack detection in IoT networks[J]. IEEE Access, 2024, 12: 42357−42368, articleTitle=null, refAbstract=null), Reference(id=1329420710080000059, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=18, rfOrder=19, authorNames=null, journalName=null, refType=null, unstructuredReference=Khan A N, Rizwan A, Ahmad R, et al. Hetero-FedIoT: A rule-based interworking architecture for heterogeneous federated IoT networks[J]. IEEE Internet of Things Journal, 2023, 11(4): 5920−5938, articleTitle=null, refAbstract=null), Reference(id=1329420710151303228, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=19, rfOrder=20, authorNames=null, journalName=null, refType=null, unstructuredReference=杨志刚, 王卓彤, 吴大鹏, 等. 物联网中带有隐私保护的鲁棒联邦学习研究[J]. 电子与信息学报, 2023, 45(12): 4235−4244, articleTitle=null, refAbstract=null), Reference(id=1329420710230995005, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=19, rfOrder=21, authorNames=null, journalName=null, refType=null, unstructuredReference=Yang Zhigang, Wang Zhuotong, Wu Dapeng, et al. Robust federated learning with privacy protection in Internet of things[J]. Journal of Electronics & Information Technology, 2023, 45(12): 4235−4244 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420710298103870, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=20, rfOrder=22, authorNames=null, journalName=null, refType=null, unstructuredReference=肖雄, 唐卓, 肖斌, 等. 联邦学习的隐私保护与安全防御研究综述[J]. 计算机学报, 2023, 46(5): 1019−1044, articleTitle=null, refAbstract=null), Reference(id=1329420710398767168, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=20, rfOrder=23, authorNames=null, journalName=null, refType=null, unstructuredReference=Xiao Xiong, Tang Zhuo, Xiao Bin, et al. A survey on privacy protection and security defense of federated learning[J]. Chinese Journal of Computers, 2023, 46(5): 1019−1044 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420710679785537, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=21, rfOrder=24, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhao Yue, Li Meng, Lai Liangzhen, et al. Federated learning with non-iid data[J]. arXiv preprint, arXiv: 1806.00582, 2018, articleTitle=null, refAbstract=null), Reference(id=1329420710889500738, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=22, rfOrder=25, authorNames=null, journalName=null, refType=null, unstructuredReference=Kairouz P, McMahan H B. Advances and open problems in federated learning[J]. Foundations and Trends in Machine Learning, 2021, 14(1-2): 1−210, articleTitle=null, refAbstract=null), Reference(id=1329420710960803907, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=23, rfOrder=26, authorNames=null, journalName=null, refType=null, unstructuredReference=穆旭彤, 程珂, 宋安霄, 等. 抗拜占庭攻击的隐私保护联邦学习[J]. 计算机学报, 2024, 47(4): 842−861, articleTitle=null, refAbstract=null), Reference(id=1329420711027912772, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=23, rfOrder=27, authorNames=null, journalName=null, refType=null, unstructuredReference=Mu Xutong, Cheng Ke, Song Anxiao, et al. Privacy-preserving federated learning against Byzantine attacks[J]. Chinese Journal of Computers, 2024, 47(4): 842−861 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420711099215941, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=24, rfOrder=28, authorNames=null, journalName=null, refType=null, unstructuredReference=Chen Weilun, Zhang Zhaoxiang, Hu Xiaolin, et al. Boosting decision-based black-box adversarial attacks with random sign flip[C]//Proc of the 16th European Conf on Computer Vision (ECCV). Cham: Springer, 2020: 276−293, articleTitle=null, refAbstract=null), Reference(id=1329420711162130502, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=25, rfOrder=29, authorNames=null, journalName=null, refType=null, unstructuredReference=Baruch G, Baruch M, Goldberg Y. A little is enough: Circumventing defenses for distributed learning[C]//Proc of the 33rd Conf on Neural Information Processing Systems (NeurIPS). New York: Curran Associates, 2019: 8632−8642, articleTitle=null, refAbstract=null), Reference(id=1329420711229239367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=26, rfOrder=30, authorNames=null, journalName=null, refType=null, unstructuredReference=Shejwalkar V, Houmansadr A. Manipulating the Byzantine: Optimizing model poisoning attacks and defenses for federated learning[C]//Proc of the 28th Annual Network and Distributed System Security Symp (NDSS). San Diego: Internet Society, 2021: 1−18, articleTitle=null, refAbstract=null), Reference(id=1329420711308931144, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=27, rfOrder=31, authorNames=null, journalName=null, refType=null, unstructuredReference=高鸿峰, 黄浩, 田有亮. 基于多方计算的安全拜占庭弹性联邦学习[J]. 通信学报, 2025, 46(2): 108−122, articleTitle=null, refAbstract=null), Reference(id=1329420711376040009, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=27, rfOrder=32, authorNames=null, journalName=null, refType=null, unstructuredReference=Gao Hongfeng, Huang Hao, Tian Youliang. Secure Byzantine-elastic federated learning based on multi-party computation[J]. Journal on Communications, 2025, 46(2): 108−122 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329420711447343178, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=28, rfOrder=33, authorNames=null, journalName=null, refType=null, unstructuredReference=Fang Minghong, Cao Xiaoyu, Jia Jinyuan, et al. Local model poisoning attacks to Byzantine-robust federated learning[C]//Proc of the 29th USENIX Security Symp (USENIX Security). Berkeley, CA: USENIX Association, 2020: 1605−1622, articleTitle=null, refAbstract=null), Reference(id=1329420711501869131, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=29, rfOrder=34, authorNames=null, journalName=null, refType=null, unstructuredReference=Allouah Y, Farhadkhani S, Guerraoui R, et al. Fixing by mixing: A recipe for optimal Byzantine ML under heterogeneity[C]//Proc of the 26th Int Conf on Artificial Intelligence and Statistics. New York: PMLR, 2023: 1232−1300, articleTitle=null, refAbstract=null), Reference(id=1329420711585755212, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=30, rfOrder=35, authorNames=null, journalName=null, refType=null, unstructuredReference=Zuo Shiyuan, Yan Xingrun, Fan Rongfei, et al. Byzantine-resilient federated learning employing normalized gradients on non-IID datasets[DB/OL]. arXiv preprint, arXiv: 2408.09539, 2024, articleTitle=null, refAbstract=null), Reference(id=1329420711652864077, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=31, rfOrder=36, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Haitao, Yao Ziming, Zhang Luyang, et al. Denial-of-service or fine-grained control: Towards flexible model poisoning attacks on federated learning[J]. arXiv preprint, arXiv: 2304.10783, 2023, articleTitle=null, refAbstract=null), Reference(id=1329420711711584334, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=32, rfOrder=37, authorNames=null, journalName=null, refType=null, unstructuredReference=Thabtah F, Hammoud S, Kamalov F, et al. Data imbalance in classification: Experimental evaluation[J]. Information Sciences, 2020, 513: 429−441, articleTitle=null, refAbstract=null), Reference(id=1329420711774498895, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=33, rfOrder=38, authorNames=null, journalName=null, refType=null, unstructuredReference=Moustafa N, Slay J. UNSW-NB15: A comprehensive data set for network intrusion detection systems (UNSW-NB15 network data set)[C]//Proc of the 2015 Military Communications and Information Systems Conf (MilCIS). Piscataway, NJ: IEEE, 2015: 1−6, articleTitle=null, refAbstract=null), Reference(id=1329420711879356496, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=34, rfOrder=39, authorNames=null, journalName=null, refType=null, unstructuredReference=Sharafaldin I, Lashkari A H, Ghorbani A A. Toward generating a new intrusion detection dataset and intrusion traffic characterization[C]//Proc of the 4th Int Conf on Information Systems Security and Privacy (ICISSP). Funchal, Portugal: SciTePress, 2018: 108−116, articleTitle=null, refAbstract=null), Reference(id=1329420712034545745, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=35, rfOrder=40, authorNames=null, journalName=null, refType=null, unstructuredReference=Garcia S, Parmisano A, Erquiaga M J. IoT−23: A labeled dataset with malicious and benign IoT network traffic (Version 1.0. 0)[DS/OL]. Zenodo, 2020. http://doi.org.10.5281/zenodo.4743746, articleTitle=null, refAbstract=null)], funds=null, companyList=[AuthorCompany(id=1329420704539325418, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=1, ext=[AuthorCompanyExt(id=1329420704572879851, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1329420704606434284, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1国防科技大学计算机学院 长沙 410073)]), AuthorCompany(id=1329420704677737453, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=2, ext=[AuthorCompanyExt(id=1329420704681931758, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2College of Network and Information, Information Support Force Engineering University, Wuhan 430030), AuthorCompanyExt(id=1329420704690320367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2信息支援部队工程大学网络信息学院 武汉 430030)]), AuthorCompany(id=1329420704749040624, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=3, ext=[AuthorCompanyExt(id=1329420704757429233, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Cyber Science and Technology, Hunan University, Changsha 410082), AuthorCompanyExt(id=1329420704765817842, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3湖南大学网络空间安全学院 长沙 410082)])], figs=[ArticleFig(id=1329420706728751120, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Figure 1, caption=Architecture diagram of GF-FedIDS method, figureFileSmall=FLe+PuTfWXQvJ5VbKY8FOg==, figureFileBig=0xJs81KT7DaZOBcIG91h5w==, tableContent=null), ArticleFig(id=1329420706787471377, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=图1, caption=GF-FedIDS方法架构图, figureFileSmall=FLe+PuTfWXQvJ5VbKY8FOg==, figureFileBig=0xJs81KT7DaZOBcIG91h5w==, tableContent=null), ArticleFig(id=1329420706955243538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Figure 2, caption=ADR heatmap across datasets, figureFileSmall=twSUB49pysDPra4RpYINgQ==, figureFileBig=M5eHdwCNwykwrRsMJrKVzw==, tableContent=null), ArticleFig(id=1329420707022352403, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=图2, caption=跨数据集ADR热力图, figureFileSmall=twSUB49pysDPra4RpYINgQ==, figureFileBig=M5eHdwCNwykwrRsMJrKVzw==, tableContent=null), ArticleFig(id=1329420707076878356, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Figure 3, caption=Comparative diagram of ADR performance across datasets, figureFileSmall=KsJE0xuwV9PT+XwZWmO7Ww==, figureFileBig=Id0mg451UZKKKswibwNrHQ==, tableContent=null), ArticleFig(id=1329420707148181525, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=图3, caption=跨数据集ADR性能对比图, figureFileSmall=KsJE0xuwV9PT+XwZWmO7Ww==, figureFileBig=Id0mg451UZKKKswibwNrHQ==, tableContent=null), ArticleFig(id=1329420707206901782, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Figure 4, caption=ADR curves with different AR settings, figureFileSmall=yL8+AxFQT3bxZGDr1aP9JQ==, figureFileBig=HFh2q5CMB6Kr6sb9bDWG3A==, tableContent=null), ArticleFig(id=1329420707320147991, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=图4, caption=不同AR设定下的ADR曲线, figureFileSmall=yL8+AxFQT3bxZGDr1aP9JQ==, figureFileBig=HFh2q5CMB6Kr6sb9bDWG3A==, tableContent=null), ArticleFig(id=1329420707399839768, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Figure 5, caption=Barplots of ADR performance under different mixed attack settings, figureFileSmall=jSLdZJOYfnnyKzVS6hxxQg==, figureFileBig=UlaLXvB1kYsvgzsJMrgusA==, tableContent=null), ArticleFig(id=1329420707471142937, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=图5, caption=混合攻击场景下ADR性能柱状图, figureFileSmall=jSLdZJOYfnnyKzVS6hxxQg==, figureFileBig=UlaLXvB1kYsvgzsJMrgusA==, tableContent=null), ArticleFig(id=1329420707542446106, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 1, caption=

Mapping of Attack Type to Optimal Defense

, figureFileSmall=null, figureFileBig=null, tableContent=
攻击类型CSNRVR最优防御
SF$ \lt {0} $$ >{1} $$ >{1} $Median
PT$ \approx {0} $$ \approx {1} $$ >{1} $Median
GS$ \approx {0} $$ >{1} $$ \gg {1} $NormClip
BT$ >{0}. {5} $$ \gg {1} $$ <{1} $TrimMean
TrimAware$ >{0}. {5} $$ \gg {1} $$ <{1} $TrimMean
LIE$ >{0}. {5} $$ \approx {1} $$ <{1} $Krum
MM$ >{0}. {5} $$ \approx {1} $$ <{1} $Krum
AD$ \approx {0} $$ >{1} $$ \gg {1} $NormClip+TrimMean
), ArticleFig(id=1329420707613749275, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表1, caption=

攻击类型与最优防御的映射

, figureFileSmall=null, figureFileBig=null, tableContent=
攻击类型CSNRVR最优防御
SF$ \lt {0} $$ >{1} $$ >{1} $Median
PT$ \approx {0} $$ \approx {1} $$ >{1} $Median
GS$ \approx {0} $$ >{1} $$ \gg {1} $NormClip
BT$ >{0}. {5} $$ \gg {1} $$ <{1} $TrimMean
TrimAware$ >{0}. {5} $$ \gg {1} $$ <{1} $TrimMean
LIE$ >{0}. {5} $$ \approx {1} $$ <{1} $Krum
MM$ >{0}. {5} $$ \approx {1} $$ <{1} $Krum
AD$ \approx {0} $$ >{1} $$ \gg {1} $NormClip+TrimMean
), ArticleFig(id=1329420707676663836, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 2, caption=

Comparsion of Different Approaches for ADR on UNSW-NB15 Dataset (AR=10%, MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
方法SFPTGSBTTrimAwareLIEMMAD平均
FedAvg0.2950.4100.2590.4620.4160.3430.4010.2060.349
Median0.6370.4380.2100.3250.2590.3820.3940.1470.349
TrimMean0.2550.4050.4330.6380.5940.3290.3450.1800.397
Krum0.5430.4880.3210.3930.3950.6100.6220.3000.459
FedProx0.3320.3980.3100.5240.4440.3740.4380.1890.376
NormClip0.2300.3630.6850.5260.4870.3000.3390.1240.381
Multi-Krum0.5540.4770.3870.5610.5220.5790.5850.2780.493
Bulyan0.5620.4430.3960.5390.5410.6210.6010.3010.501
GF-FedIDS0.6450.5200.6830.6430.6270.6440.6370.3610.595
基线平均0.4260.4280.3740.4960.4570.4420.4660.2160.413
基线最佳0.6370.4880.6750.6380.5940.6210.6220.3010.572
最佳性能比1.011.071.011.011.061.041.021.201.04
), ArticleFig(id=1329420707752161309, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表2, caption=

UNSW-NB15数据集上各方法ADR对比(AR=10%,MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
方法SFPTGSBTTrimAwareLIEMMAD平均
FedAvg0.2950.4100.2590.4620.4160.3430.4010.2060.349
Median0.6370.4380.2100.3250.2590.3820.3940.1470.349
TrimMean0.2550.4050.4330.6380.5940.3290.3450.1800.397
Krum0.5430.4880.3210.3930.3950.6100.6220.3000.459
FedProx0.3320.3980.3100.5240.4440.3740.4380.1890.376
NormClip0.2300.3630.6850.5260.4870.3000.3390.1240.381
Multi-Krum0.5540.4770.3870.5610.5220.5790.5850.2780.493
Bulyan0.5620.4430.3960.5390.5410.6210.6010.3010.501
GF-FedIDS0.6450.5200.6830.6430.6270.6440.6370.3610.595
基线平均0.4260.4280.3740.4960.4570.4420.4660.2160.413
基线最佳0.6370.4880.6750.6380.5940.6210.6220.3010.572
最佳性能比1.011.071.011.011.061.041.021.201.04
), ArticleFig(id=1329420707823464478, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 3, caption=

Comparison of Different Approaches for ADR Under Different AR Proportion (MCR=20%, Average on Eight Types of Attacks)

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集方法AR
5%10%15%20%30%50%
UNSW-
NB15
基线平均0.1780.2310.3050.3820.4890.743
基线最佳0.2170.2480.3420.4040.5230.791
GF-FedIDS0.4480.4320.4350.4940.5750.753
最佳性能比2.061.741.271.221.100.95
CICIDS-
2017
基线平均0.1770.2540.3160.3750.5110.774
基线最佳0.2020.2780.3570.4060.5400.820
GF-FedIDS0.4630.4470.4610.5070.5930.803
最佳性能比2.291.611.291.251.100.98
CIC-IoT-
2023
基线平均0.1700.2310.2950.3610.4620.726
基线最佳0.1970.2550.3110.3850.4980.779
GF-FedIDS0.4140.4160.4340.4680.5560.740
最佳性能比2.101.631.391.221.120.95
), ArticleFig(id=1329420707903156255, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表3, caption=

不同AR占比下各方法ADR对比(MCR=20%,8种攻击平均)

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集方法AR
5%10%15%20%30%50%
UNSW-
NB15
基线平均0.1780.2310.3050.3820.4890.743
基线最佳0.2170.2480.3420.4040.5230.791
GF-FedIDS0.4480.4320.4350.4940.5750.753
最佳性能比2.061.741.271.221.100.95
CICIDS-
2017
基线平均0.1770.2540.3160.3750.5110.774
基线最佳0.2020.2780.3570.4060.5400.820
GF-FedIDS0.4630.4470.4610.5070.5930.803
最佳性能比2.291.611.291.251.100.98
CIC-IoT-
2023
基线平均0.1700.2310.2950.3610.4620.726
基线最佳0.1970.2550.3110.3850.4980.779
GF-FedIDS0.4140.4160.4340.4680.5560.740
最佳性能比2.101.631.391.221.120.95
), ArticleFig(id=1329420707961876512, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 4, caption=

Comparison of ADR Under Different Mixed Attack Settings (AR=10%,MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集方法SF+LIEGS+BT8种随
机混合
4种自适
应混合
平均
UNSW-
NB15
基线平均0.2800.2770.1820.1780.229
基线最佳0.4210.3630.2760.2540.329
GF-FedIDS0.6110.6500.5570.5310.587
最佳性能比1.451.792.022.091.78
CICIDS-
2017
基线平均0.2870.2870.1880.1840.237
基线最佳0.4420.3820.2800.2640.342
GF-FedIDS0.6390.6570.5800.5460.605
最佳性能比1.451.722.072.071.77
CIC-IoT-
2023
基线平均0.2570.2610.1680.1670.213
基线最佳0.3920.3540.2580.2360.310
GF-FedIDS0.5960.6050.5410.5040.562
最佳性能比1.521.712.092.141.81
), ArticleFig(id=1329420708028985377, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表4, caption=

混合攻击场景下ADR对比(AR=10%,MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集方法SF+LIEGS+BT8种随
机混合
4种自适
应混合
平均
UNSW-
NB15
基线平均0.2800.2770.1820.1780.229
基线最佳0.4210.3630.2760.2540.329
GF-FedIDS0.6110.6500.5570.5310.587
最佳性能比1.451.792.022.091.78
CICIDS-
2017
基线平均0.2870.2870.1880.1840.237
基线最佳0.4420.3820.2800.2640.342
GF-FedIDS0.6390.6570.5800.5460.605
最佳性能比1.451.722.072.071.77
CIC-IoT-
2023
基线平均0.2570.2610.1680.1670.213
基线最佳0.3920.3540.2580.2360.310
GF-FedIDS0.5960.6050.5410.5040.562
最佳性能比1.521.712.092.141.81
), ArticleFig(id=1329420708112871458, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 5, caption=

Results of Ablation study on UNSW-NB15 Dataset (AR=10%,MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
配置ADRACCFPR
已知攻击类型选择最优固定基线策略0.5380.830.048
GF-FedIDS(无IW)0.5150.860.032
GF-FedIDS(完整方法)0.5950.870.028
), ArticleFig(id=1329420708184174627, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表5, caption=

在UNSW-NB15数据集上的消融实验结果(AR=10%,MCR=20%)

, figureFileSmall=null, figureFileBig=null, tableContent=
配置ADRACCFPR
已知攻击类型选择最优固定基线策略0.5380.830.048
GF-FedIDS(无IW)0.5150.860.032
GF-FedIDS(完整方法)0.5950.870.028
), ArticleFig(id=1329420708242894884, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=EN, label=Table 6, caption=

Comparison of Computing Time Consumption of Different approaches

, figureFileSmall=null, figureFileBig=null, tableContent=
方法时间复杂度每轮耗时/s开销比
FedAvg$ {O}\left({n}{d}\right) $0.451.00
Median$ {O}\left({n}{d}\log {n}\right) $0.481.07
Krum$ {O}\left({{n}}^{{2}}{d}\right) $0.821.82
Multi-Krum$ {O}\left({{n}}^{{2}}{d}\right) $1.854.11
Bulyan$ {O}\left({{n}}^{{2}}{d}+{n}{m}\log {m}\right) $2.204.89
GF-FedIDS$ {O}\left({{n}}^{{2}}{d}+{n}{m}\log {m}\right) $0.521.16
), ArticleFig(id=1329420708297420837, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, language=CN, label=表6, caption=

各方法计算时间开销对比

, figureFileSmall=null, figureFileBig=null, tableContent=
方法时间复杂度每轮耗时/s开销比
FedAvg$ {O}\left({n}{d}\right) $0.451.00
Median$ {O}\left({n}{d}\log {n}\right) $0.481.07
Krum$ {O}\left({{n}}^{{2}}{d}\right) $0.821.82
Multi-Krum$ {O}\left({{n}}^{{2}}{d}\right) $1.854.11
Bulyan$ {O}\left({{n}}^{{2}}{d}+{n}{m}\log {m}\right) $2.204.89
GF-FedIDS$ {O}\left({{n}}^{{2}}{d}+{n}{m}\log {m}\right) $0.521.16
)], attaches=null, journal=Journal(id=1324759606947315908, delFlag=0, nameCn=计算机研究与发展, nameEn=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, issn=1000-1239, eissn=null, cn=11-1777/TP, coden=null, periodic=0, language=CN, oaType=null, ccby=null, superviseOffice=null, ownerOffice=null, pubOffice=null, editorOffice=null, officeType=null, aims=null, clcCode=null, officeProv=null, officeCity=null, officeAddr=null, officeZip=null, officeEmail=null, officePhone=null, editDirector=null, officeDirector=null, officeDirectorPhone=null, officeStaffNum=null, officeEmpNum=null, coverPicUrl=k2l8zgqwstuxxevFX2qxVA==, journalPrice=null, startedYear=null, abbrevIsoEn=Journal of Computer Research and Development, journalRemark=null, publicationField=null, createdTime=1793776903141, updatedTime=1793850698612, createdBy=18614031015, updatedBy=13701087609, firstLetterCn=J, firstLetterEn=J, subjectCode=Electronic Information, subjectName=null, subjectCodeEn=Electronic Information, subjectNameEn=null, picCn=k2l8zgqwstuxxevFX2qxVA==, picEn=r6k0Ry+RD+c/VXj5QarmDw==, jcr=null, cjcr=null, exts=[JournalExt(id=1325069127710286688, language=CN, name=计算机研究与发展, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698640, updatedTime=1793850698640, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""}), JournalExt(id=1325069127777395553, language=EN, name=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698656, updatedTime=1793850698656, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""})], databaseList=null, tenantJournalId=1324759914146529478, websiteList=[Website(id=1325070300139892771, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN, language=CN, createTime=1793850978169, createBy=18614031015, updateTime=1793851040259, updateBy=18614031015, name=计算机研究与发展-中文, tplId=1146099689490845704, title=计算机研究与发展, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070893751349360, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=articleTextType, value=kx, createTime=1793851119697, updateTime=1793851119697, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893730377837, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=banner, value=null, createTime=1793851119692, updateTime=1793851119692, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893772320883, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=grayFlag, value=0, createTime=1793851119702, updateTime=1793851119702, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893721989228, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851119690, updateTime=1793851119690, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893784903797, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=minRunFlag, value=0, createTime=1793851119705, updateTime=1793851119705, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893742960751, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic, createTime=1793851119695, updateTime=1793851119695, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893780709492, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=silenceFlag, value=0, createTime=1793851119704, updateTime=1793851119704, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893738766446, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1793851119694, updateTime=1793851119694, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893755543665, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeColor, value=null, createTime=1793851119698, updateTime=1793851119698, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893763932274, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeStyle, value=null, createTime=1793851119700, updateTime=1793851119700, creator=18614031015, updator=18614031015)]), Website(id=1325070300580294698, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN, language=EN, createTime=1793850978274, createBy=18614031015, updateTime=1793851036638, updateBy=18614031015, name=计算机研究与发展-英文, tplId=1146101810881728533, title=Journal of Computer Research and Development, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070925288321148, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=articleTextType, value=kx, createTime=1793851127216, updateTime=1793851127216, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925263155321, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=banner, value=null, createTime=1793851127210, updateTime=1793851127210, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925309292671, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=grayFlag, value=0, createTime=1793851127221, updateTime=1793851127221, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925254766712, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851127208, updateTime=1793851127208, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925326069889, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=minRunFlag, value=0, createTime=1793851127225, updateTime=1793851127225, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925279932539, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic, createTime=1793851127214, updateTime=1793851127214, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925317681280, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=silenceFlag, value=0, createTime=1793851127223, updateTime=1793851127223, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925271543930, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_en_623/, createTime=1793851127212, updateTime=1793851127212, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925292515453, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeColor, value=null, createTime=1793851127217, updateTime=1793851127217, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925300904062, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeStyle, value=null, createTime=1793851127219, updateTime=1793851127219, creator=18614031015, updator=18614031015)])], journalTitle=计算机研究与发展, weixinUrl=null, journalUrl=https://crad.ict.ac.cn/, iacademicId=null, status=1, seqNo=null, journalTitleEn=Journal of Computer Research and Development, journalPhotoCn=k2l8zgqwstuxxevFX2qxVA==, journalPhotoEn=r6k0Ry+RD+c/VXj5QarmDw==, journalFirstLetter=J, journalRecommend=null, journalNew=null, journalCollection=null, jcrJf=null, cjcrJf=null, jcrJfStr=null, cjcrJfStr=null, submissionFirstDecision=null, sciSubjectClassification=null, casSubjectClassification=null, citeScore=null, totalCitationFrequency=null, icpCode=null, psCode=null, advertisingLicenseCode=null, copyrightInformation=null, country=null, option=, provinceCode=null, provinceName=null, collectFlag=false, interPubPlatform=, interPubPlatformUrl=null), detailUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202660484, detailUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/10.7544/issn1000-1239.202660484, pdfUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/PDF/10.7544/issn1000-1239.202660484, pdfUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/PDF/10.7544/issn1000-1239.202660484, aliStartDate=0, aliEndDate=0, collectionFlag=false, citedCount=null, citedUrl=null, previewStatus=0, delFlag=0, hasFullText=1, orderTime=1790784000000, fullTextJson=null, articleText=null, reference=null)
收藏切换
基于梯度指纹决策的拜占庭鲁棒联邦学习网络入侵检测方法
姜植元 1 , 杨刚 2, * , 王勇军 1 , 秦拯 3
计算机研究与发展 | 动态异构网络中海量异质终端安全防护技术专题 2026,63(10): 2576-2595
收起
收藏切换
计算机研究与发展 |动态异构网络中海量异质终端安全防护技术专题 2026 , 63 (10) : 2576 -2595
基于梯度指纹决策的拜占庭鲁棒联邦学习网络入侵检测方法
全屏
[Author(id=1329420704824538100, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=jzy@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420704904229878, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420704824538100, language=EN, stringName=Zhiyuan Jiang, firstName=Zhiyuan, middleName=null, lastName=Jiang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

"}, bioImg=null, bioContent=

born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420704967144439, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420704824538100, language=CN, stringName=姜植元, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1国防科技大学计算机学院 长沙 410073, bio={"img":"ArsWPHysmqF0+Pm30DbQhA==","content":"

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

"}, bioImg=ArsWPHysmqF0+Pm30DbQhA==, bioContent=

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704539325418, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=1, ext=[AuthorCompanyExt(id=1329420704572879851, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1329420704606434284, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1国防科技大学计算机学院 长沙 410073)])]), Author(id=1329420705025864697, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=yanggang11@nudt.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329420705113945083, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705025864697, language=EN, stringName=Gang Yang, firstName=Gang, middleName=null, lastName=Yang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, *, address=2College of Network and Information, Information Support Force Engineering University, Wuhan 430030, bio={"content":"

born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

"}, bioImg=null, bioContent=

born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705189442556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705025864697, language=CN, stringName=杨刚, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, *, address=2信息支援部队工程大学网络信息学院 武汉 430030, bio={"img":"yHBJbn6/ehSweB35Ac677Q==","content":"

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

"}, bioImg=yHBJbn6/ehSweB35Ac677Q==, bioContent=

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704677737453, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=2, ext=[AuthorCompanyExt(id=1329420704681931758, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2College of Network and Information, Information Support Force Engineering University, Wuhan 430030), AuthorCompanyExt(id=1329420704690320367, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704677737453, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2信息支援部队工程大学网络信息学院 武汉 430030)])]), Author(id=1329420705264940030, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420705373991936, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705264940030, language=EN, stringName=Yongjun Wang, firstName=Yongjun, middleName=null, lastName=Wang, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, bio={"content":"

born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

"}, bioImg=null, bioContent=

born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705445294080, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705264940030, language=CN, stringName=王勇军, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1国防科技大学计算机学院 长沙 410073, bio={"img":"HgQj/LSlTautaN/aYy12kg==","content":"

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

"}, bioImg=HgQj/LSlTautaN/aYy12kg==, bioContent=

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704539325418, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=1, ext=[AuthorCompanyExt(id=1329420704572879851, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073), AuthorCompanyExt(id=1329420704606434284, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704539325418, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1国防科技大学计算机学院 长沙 410073)])]), Author(id=1329420705508208642, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329420705579511812, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705508208642, language=EN, stringName=Zheng Qin, firstName=Zheng, middleName=null, lastName=Qin, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3College of Cyber Science and Technology, Hunan University, Changsha 410082, bio={"content":"

born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

"}, bioImg=null, bioContent=

born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329420705663397893, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, authorId=1329420705508208642, language=CN, stringName=秦拯, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3湖南大学网络空间安全学院 长沙 410082, bio={"img":"Y68ndMp8waECsNVk3tfqpQ==","content":"

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

"}, bioImg=Y68ndMp8waECsNVk3tfqpQ==, bioContent=

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329420704749040624, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, xref=3, ext=[AuthorCompanyExt(id=1329420704757429233, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Cyber Science and Technology, Hunan University, Changsha 410082), AuthorCompanyExt(id=1329420704765817842, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329420702152766421, companyId=1329420704749040624, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3湖南大学网络空间安全学院 长沙 410082)])])]
姜植元1 , 杨刚2, * , 王勇军1, 秦拯3
作者信息
  • 1国防科技大学计算机学院 长沙 410073
  • 2信息支援部队工程大学网络信息学院 武汉 430030
  • 3湖南大学网络空间安全学院 长沙 410082
通讯作者:
作者简介:

作者贡献声明

姜植元提出了研究方法与解决思路,以及方法论的整体构想;杨刚负责具体设计实现并撰写论文;王勇军、秦拯为论文提出指导意见。

1994年生。副教授,硕士生导师。主要研究方向为软件安全、系统安全

1993年生。讲师,硕士生导师。主要研究方向为软件安全、流量分析

1971年生。教授,博士生导师。主要研究方向为信息安全、网络安全

1969年生。教授,博士生导师。主要研究方向为网络与信息安全、人工智能安全、知识工程

Byzantine-Robust Federated Learning for Network Intrusion Detection via Gradient Fingerprint-Driven Decision
Zhiyuan Jiang1 , Gang Yang2, * , Yongjun Wang1, Zheng Qin3
Affiliations
  • 1College of Computer Science and Technology, National University of Defense Technology, Changsha 410073
  • 2College of Network and Information, Information Support Force Engineering University, Wuhan 430030
  • 3College of Cyber Science and Technology, Hunan University, Changsha 410082
  • born in 1994. Associate professor, master supervisor. His main research interests include software security and system security

    born in 1993. Lecturer, master supervisor. His main research interests include software security and network traffic analysis

    born in 1971. Professor, PhD supervisor. His main research interests include information security and network security

    born in 1969. Professor, PhD supervisor. His main research interests include network and information security, artificial intelligence security, and knowledge engineering

出版时间: 2026-10-01 doi: 10.7544/issn1000-1239.202660484
文章导航
收藏切换

动态异构网络中,海量异质终端导致安全数据分散异构,联邦学习虽能保护数据隐私实现跨终端的协同入侵检测,但其开放架构易受拜占庭攻击,即恶意客户端通过上传恶意梯度破坏全局模型。现有拜占庭鲁棒方法通常仅在特定攻击类型上有效,难以覆盖真实场景中多类型攻击共存的态势,且在攻击样本稀疏的真实流量环境下易产生大量漏报,多防御堆叠的计算开销亦难以适配资源受限终端。对此,提出基于梯度指纹决策的拜占庭鲁棒联邦学习入侵检测框架GF-FedIDS(gradient fingerprint-driven federated intrusion detection system),通过对典型拜占庭攻击梯度统计特性的分析,建立攻击模式到最优防御的确定性映射,揭示其因果关联;设计梯度指纹决策树,实现自适应选择单一最优防御策略,以单位计算开销达成全覆盖鲁棒聚合;引入类别感知实例加权机制,缓解数据不平衡下的漏报问题。在3个公开数据集、8种攻击类型条件下的广泛实验表明,在单攻击场景下,GF-FedIDS的攻击检测率(ADR)在跨数据集场景下普遍略优于最优基线;在混合攻击场景下,ADR为最优基线的1.45~2.14倍;在攻击率为5%的极端不平衡条件下,ADR为最优基线的2.06~2.29倍。此外,GF-FedIDS与FedAvg相比仅增加16%的计算开销,优于现有集成防御方法。

联邦学习  /  拜占庭攻击  /  入侵检测  /  梯度指纹  /  非独立同分布数据

In dynamic heterogeneous networks, massive heterogeneous terminals result in dispersed and diverse security data. While federated learning enables collaborative intrusion detection across terminals without compromising data privacy, its open architecture is inherently vulnerable to Byzantine attacks, where malicious clients poison the global model by uploading crafted gradients. Existing Byzantine-robust methods are typically effective only against specific attack types and struggle to cope with the coexistence of multiple attacks in real-world environments. Moreover, they suffer from high false-negative rates under sparse real-traffic conditions, and the computational overhead of stacking multiple defenses is prohibitive for resource-constrained terminals. To address these challenges, we propose GF-FedIDS, a Byzantine-robust federated learning intrusion detection framework based on gradient fingerprint decision-making. By analyzing the gradient statistics of typical Byzantine attacks, we establish a deterministic mapping from attack patterns to optimal defenses and reveal their causal relationship. A gradient fingerprint decision tree is designed to adaptively select a single optimal defense strategy, achieving full-coverage robust aggregation at unit computational cost. Furthermore, a class-aware instance weighting mechanism is introduced to mitigate missed detections under data imbalance. Extensive experiments on three public datasets and eight attack types demonstrate that, in single-attack scenarios, the attack detection rate (ADR) of GF-FedIDS outperforms that of the best baseline across three datasets; in mixed-attack scenarios, ADR reaches 1.45–2.14 times; and under extreme imbalance with a 5% attack rate, ADR is 2.06–2.29 times that of the best baseline. Additionally, GF-FedIDS incurs only a 16% computational overhead compared with FedAvg, outperforming existing ensemble defense methods.

federated learning  /  Byzantine attack  /  intrusion detection  /  gradient fingerprint  /  Non-IID data
姜植元, 杨刚, 王勇军, 秦拯. 基于梯度指纹决策的拜占庭鲁棒联邦学习网络入侵检测方法. 计算机研究与发展, 2026 , 63 (10) : 2576 -2595 . DOI: 10.7544/issn1000-1239.202660484
Zhiyuan Jiang, Gang Yang, Yongjun Wang, Zheng Qin. Byzantine-Robust Federated Learning for Network Intrusion Detection via Gradient Fingerprint-Driven Decision[J]. Journal of Computer Research and Development, 2026 , 63 (10) : 2576 -2595 . DOI: 10.7544/issn1000-1239.202660484
随着物联网、云边端协同计算与下一代通信技术的快速演进,动态异构网络已成为互联网组网的主流形态,海量异质终端的规模化部署成为常态[1]。这使网络入侵检测面临数据分散、隐私敏感、标注稀缺三重困境。传统集中式检测要求将终端流量数据统一汇聚至云端分析,不仅面临高昂的带宽开销与实时性瓶颈,更因原始网络流量可能包含用户通信内容、设备指纹等敏感信息,难以满足日益严苛的数据隐私合规要求。联邦学习作为一种“数据不动模型动”的分布式机器学习范式,为破解上述困境提供了新的技术路径[2],异质终端在本地利用私有流量数据训练检测模型,仅向中心服务器上传模型参数,在保护数据隐私的前提下实现跨终端入侵检测能力的协同提升[3-5]。
然而,联邦学习的开放协作架构在另一方面引入了新的安全问题。由于模型参数在终端与服务器之间公开传输,攻击者可采用拜占庭攻击,通过攻陷部分终端、上传精心构造的恶意梯度,干扰全局模型的聚合方向,使检测能力严重退化甚至完全失效,这给联邦学习落地于安全敏感场景带来重要挑战。更为严峻的是,动态异构网络中的终端天然呈现海量异质特征——不同终端的硬件能力、网络环境、通信模式和使用场景差异显著,导致各客户端本地流量数据呈典型的非独立同分布(non-independent and identically distributed,Non-IID)。这种数据异质性使得良性客户端之间的本地模型更新本就存在明显偏差,恶意更新可借由这种“正常偏差”隐藏自身,大幅增加了拜占庭攻击的辨识难度[6-7]。与此同时,在现实网络安全防御场景下,攻击样本在真实网络流量中通常占比很低,其极端类别不平衡性使得模型在线联邦学习训练过程被正常样本主导,攻击类别的特征难以被充分学习,导致模型优化目标的天然偏置,产生严重的攻击漏检风险,大大削弱了入侵检测系统的可靠性。
为应对拜占庭攻击,近年来Krum[8]、坐标中位数(Median)[8]、修剪均值(TrimMean)[9]、范数裁剪(NormClip)[10]、FedProx[11]等一系列鲁棒聚合方法被提出。这些方法通过剔除或抑制统计异常梯度,在一定程度上保障了联邦学习训练的稳定性。但现有方法在设计之初多面向通用机器学习场景下单类型拜占庭攻击,默认各类样本均衡分布、客户端数据同质,并未充分考虑网络入侵检测任务的领域特性,导致在攻击样本稀疏、数据高度异构的真实流量环境下防御效果大打折扣[12]。具体而言,不同拜占庭攻击对梯度统计特性施加了差异化扰动,不存在普适性的最优固定防御,单一防御方法通常仅在特定攻击类型上表现有效,难以应对混杂海量异构终端的动态异构网络场景中多种攻击共存的复杂态势;而单纯的多防御堆叠虽能扩大覆盖范围,其计算开销却难以适配资源受限的物联网终端。更为本质的是,上述方法均面临一个结构性悖论:鲁棒聚合操作在抑制拜占庭扰动时,不可避免地会削弱本就微弱的攻击类梯度信号,导致防御强度与攻击检测率之间存在此消彼长的困境。
针对上述问题,一种改进思路是根据当前攻击态势动态选择最合适的防御方法。近年来已有研究探索自适应防御方法,如Multi-Krum[8]、Bulyan[13],即不再依赖固定防御策略,而是根据观测自适应切换防御。然而,现有自适应方法普遍存在2个瓶颈:其一,方法选择多依赖启发式规则或经验阈值,缺乏从攻击模式到最优防御的严格理论映射;其二,未考虑类别不平衡与拜占庭防御的耦合效应,无法避免防御操作对攻击类梯度信号的二次压制。
鉴于上述考虑,本文的核心思路在于利用不同拜占庭攻击对梯度的统计特性施加了差异化、可辨识的确定性扰动,如同在梯度空间中留下独特的“指纹”。基于这一洞察,我们提出基于梯度指纹决策的拜占庭鲁棒联邦学习入侵检测框架GF-FedIDS(gradient fingerprint-driven federated intrusion detection system)。GF-FedIDS由3个关键组件组成:梯度指纹提取、指纹决策树自适应防御和类别感知实例加权。首先,对典型拜占庭攻击的梯度统计特性进行系统分析,提取方向余弦、范数比、方差比构成三维梯度指纹,建立攻击模式到最优防御方法的确定性映射。随后,利用不同梯度指纹在特征空间中的可分性,构建轻量级决策树,每轮聚合前服务器提取当前梯度指纹,沿决策树路径快速判定攻击类型,自适应选择单一最优防御。该机制以单位计算开销实现多类型攻击的全覆盖鲁棒聚合,且防御选择具有可解释的理论依据。为进一步缓解极端不平衡下的漏报问题,引入类别感知实例加权机制,通过赋予攻击样本更高的优化权重,从源头补偿其在损失函数中被淹没的梯度贡献。
本文的贡献主要包括4个方面:
1)提出了基于梯度指纹的防御策略路由理论。该理论通过对典型拜占庭攻击梯度统计特性的形式化分析,提取方向余弦、范数比、方差比构成三维梯度指纹,建立了攻击模式到最优防御方法之间的因果映射关系,为自适应防御选择提供了数学基础;
2)提出了类别感知实例加权机制。针对实践网络流量中攻击样本极度稀疏的领域特性,在本地训练阶段为攻击样本赋予更高的优化权重,在训练损失函数中重新平衡正负样本的优化比重,有效解决极端不平衡场景下的漏报问题;
3)提出了基于梯度指纹决策的自适应防御方法GF-FedIDS。该框架构建轻量级指纹决策树,每轮基于提取的梯度指纹自适应选择单一最优防御策略,以单位计算开销实现多类型拜占庭攻击的全覆盖鲁棒聚合,实现“攻击识别—类别感知—自适应防御”三大机制的协同集成,突破了现有方法仅针对特定攻击有效的局限;
4)在3种公开数据集上围绕8种攻击类型条件下进行了200余次独立实验。结果表明,在单攻击场景下,GF-FedIDS的攻击检测率(ADR)普遍略优于最优基线;在混合攻击场景下,ADR为最优基线的1.45~2.14倍;在攻击率为5%的极端不平衡条件下,ADR为最优基线的2.06~2.29倍。此外,GF-FedIDS与FedAvg相比仅增加16%的计算开销,优于现有集成防御方法。
联邦学习由Brendan等人[2]于2017年系统提出,其核心思想是通过“数据不动模型动”的协同训练模式,使分布式客户端在不共享本地原始数据的前提下共同训练全局模型。该范式避免了原始数据的集中化存储与传输,在保护数据隐私的同时为海量异质终端的协同学习提供了可行的技术路径[14-15]。
近年来,联邦学习在物联网入侵检测领域取得了显著进展。Zhang等人[16]系统综述了联邦深度学习在物联网安全中的应用,指出该领域面临的通信瓶颈、数据异构与安全攻击等核心挑战。Alhasawi等人[17]利用联邦学习实现了跨多个IoT网络的协同DDoS攻击检测;Khan等人[18]提出了一种基于规则的互通架构Hetero-FedIoT,通过中间件层实现异构物联网设备间跨联邦学习框架的协同训练,解决了不同联邦学习协议和网络配置下设备的互操作问题。上述工作表明,联邦学习为异构终端构成的动态网络场景下的恶意流量检测提供了有效的技术基础[19-20]。
然而,上述方法均建立在FedAvg聚合算法[2]之上,该算法仅在客户端数据满足独立同分布(IID)假设时方能取得接近集中式训练的收敛效果[2]。在实际动态异构网络中,各客户端本地数据呈显著的Non-IID特性,FedAvg在此条件下收敛速度下降、模型性能退化已成为共识[21]。更为严峻的是,联邦学习的开放架构使其天然暴露于拜占庭攻击威胁之下。当Non-IID数据分布与拜占庭攻击相互叠加时,2种效应形成风险进一步耦合:良性客户端之间的本地更新本身存在的偏差,会增加攻击辨识难度;而拜占庭攻击则进一步放大Non-IID引入的偏差,使得全局模型偏离最优方向。这使得标准联邦学习框架在海量异质终端的入侵检测任务中面临严重的安全与性能危机,如何在此场景下实现拜占庭鲁棒的联邦学习入侵检测,成为亟待解决的关键问题。
拜占庭攻击指在联邦学习分布式训练过程中,恶意客户端通过上传精心构造的恶意更新破坏全局模型收敛的行为[22]。现有拜占庭攻击可分为数据投毒与模型投毒两大类:前者通过篡改本地训练数据间接影响梯度更新,典型方法包括符合翻转攻击和后门攻击;后者直接构造恶意梯度上传,具有更强的攻击效果[23]。Chen等人[24]系统研究了联邦学习中的拜占庭攻击,提出了SignFlip符号翻转和随机噪声等基础攻击。后续研究发展了更为复杂的变体:Baruch等人[25]提出的LIE攻击通过统计模仿使恶意梯度难以区分;Shejwalkar等人[26]提出的Min-Max和Min-Sum攻击通过约束优化最大化攻击影响;高鸿峰等人[27]与Fang等人[28]分别系统研究了本地模型投毒攻击对拜占庭鲁棒联邦学习的威胁。Allouah等人[29]从异构性角度分析了拜占庭机器学习的理论极限,提出了Fixing by mixing策略。
针对拜占庭攻击,研究者提出了多种防御方法,可分为以下3类:
1)基于稳健统计的方法。Median[8]使用各参数维度坐标中位数替代加权平均,TrimMean[9]在每个坐标维度上去除极值后取平均。此类方法实现简单,但对自适应攻击和定向攻击的防御效果有限。
2)基于正则化的方法。FedProx[11]在本地训练目标函数中添加近端项,限制本地更新与全局模型的偏离程度,从而抑制恶意更新的影响。但该方法引入额外的超参数,且对强攻击的防御效果不够理想。
3)基于梯度过滤的方法。NormClip[10]根据梯度范数裁剪异常更新,Fed-NGA[30]基于梯度对齐性进行过滤。此类方法对范数异常攻击有效,但对统计模仿类攻击的检测能力不足。
上述防御方法的共同局限在于:均采用单一固定策略,无法覆盖所有攻击类型。不同防御方法在不同攻击场景下表现差异显著:Median和TrimMean对符号翻转攻击具有天然的鲁棒性,但在自适应攻击[31]等定向攻击下近乎失效;Krum对LIE和Min-Max等统计模仿类攻击效果良好,却难以抵御自适应攻击;基于历史梯度一致性的Fed-NGA虽对多数攻击表现稳定,在定向攻击下仍部分有效;FedProx由于仅通过近端约束间接限制恶意更新幅度,在各类攻击下均仅能提供有限防护。在此基础上,研究者探索了2种扩展思路:其一,Blanchard等人[8]在提出Krum时同时提出了Multi-Krum,即选择多个评分最高的梯度进行聚合,但本质上是同一防御方法的多次应用,并未实现不同防御策略的优势互补;其二,Guerraoui 等人[13]提出的Bulyan结合了Krum与TrimMean,先以Krum筛选候选梯度,再以TrimMean聚合,属于2种防御的串联集成。然而,Bulyan虽扩大了防御覆盖范围,其串行执行方式仍无法动态适配攻击类型的变化,且2种防御的计算开销叠加,限制了其在资源受限场景下的部署效率。
形式化地,考虑由N个客户端和1个中央服务器组成的横向联邦学习系统。设第$ i $个客户端持有本地私有数据集$ {\mathcal{D}}_{i} =\{({\boldsymbol x}_{j},{y}_{j})\}_{j=1}^{{n}_{i}} $,其中$ {\boldsymbol x}_{j}\in \mathcal{X} $为输入特征(网络流量的统计特征向量),$ {y}_{j}\in \mathcal{Y}=\{0,1\} $为标签(0表示正常流量,1表示攻击流量),$ {n}_{i}=\left| {\mathcal{D}}_{i}\right| $为客户端$ i $的样本量。全局目标函数定义为所有客户端本地经验风险的最小化:
$ \underset{\boldsymbol w\in {\mathbb{R}}^{d}}{\min }F\left(\boldsymbol w\right)=\displaystyle\sum\limits_{i=1}^{N}\dfrac{{n}_{i}}{n}{F}_{i}\left(\boldsymbol w\right) \;{,} $
其中$ {F}_{i}\left(\boldsymbol w\right)=\dfrac{1}{{n}_{i}}\displaystyle\sum\limits_{\left(\boldsymbol x,\boldsymbol y\right)\in {\mathcal{D}}_{i}}\ell\left(\boldsymbol w;\boldsymbol x,\boldsymbol y\right) $为第$ i $个客户端的本地经验风险,$ n=\displaystyle\sum\limits_{i=1}^{N}{n}_{i} $为总样本量,$ \ell(\cdot ) $为损失函数(本文采用二元交叉熵损失),$ \boldsymbol w\in {\mathbb{R}}^{d} $为模型参数向量。联邦学习的训练过程通过多轮迭代进行,在第 $ t $ 轮通信中:首先,服务器将当前全局模型 $ {\boldsymbol w}^{t} $ 广播给选定的客户端集合 $ {\mathcal{S}}^{t}\subseteq \{h_1,h_2…,h_N\} $;接着,每个客户端 $ h_i\in {\mathcal{S}}^{t} $ 在本地执行 $ E $ 轮小批量随机梯度下降(SGD),计算本地模型更新$ \boldsymbol w_{i}^{t+1}={\boldsymbol w}^{t}-\eta \nabla {F}_{i}\left({\boldsymbol w}^{t}\right) $,并将更新差值(伪梯度)$ \boldsymbol g_{i}^{t}={\boldsymbol w}^{t}-\boldsymbol w_{i}^{t+1} $上传至服务器;最终,服务器执行聚合操作$ {\boldsymbol w}^{t+1} ={\boldsymbol w}^{t} -{\eta }_{\rm global}\cdot {\cal A} {\mathscr{g g}} (\{\boldsymbol g_{i}^{t}{\}}_{i\in {{\mathcal{S}}^{t}}} ) $,其中$ {\cal A} {\mathscr{g g}} (\cdot ) $为聚合函数。标准联邦学习(FedAvg[2])采用加权平均聚合,$ {\cal A} {\mathscr{g g}}(\cdot ) $取加权平均,即:
$ {\boldsymbol w}^{t+1}=\displaystyle\sum\limits_{i\in {\mathcal{S}}^{t}}\dfrac{{n}_{i}}{\displaystyle\sum\limits_{j}{n}_{j}}\boldsymbol w_{i}^{t+1} 。 $
海量异质终端场景的网络入侵检测具有区别于一般联邦学习任务的特性,这决定了其拜占庭防御策略的设计方向:1)高维稀疏性。网络流量特征通常维度达数十至上百维,且大量特征值为0。以CICIDS-2017数据集为例,78维特征中平均约40%的取值为0,这种稀疏结构使得梯度在参数空间中呈现高度不对称分布。2)极度不平衡性。正常流量通常占绝大部分,攻击流量在数量和质量上均有显著缺陷,导致传统联邦学习的优化目标易被正常样本主导,攻击样本的梯度贡献被大幅削弱。3)分布异质性。不同终端的设备类型、网络协议栈及使用模式差异显著,各本地数据分布之间差异巨大。此特性正是Non-IID与拜占庭攻击危险耦合的根源所在:良性客户端之间的本地更新本就存在系统性偏差,恶意更新可借此“正常偏移”隐蔽自身。4)漏报敏感性。安全检测场景漏检攻击的代价远高于误报正常流量,要求模型保持高召回率。
上述域特性引发2个核心矛盾:其一,Non-IID分布使各客户端的本地最优解存在系统性偏移,拜占庭攻击可借此掩盖恶意行为;其二,类别极度不平衡使攻击样本在全局聚合中自然被淹没,模型可能因优化目标偏差而漏检攻击。
在海量异质终端场景下,终端设备数量庞大、类型多样、安全等级参差不齐,恶意客户端的渗透风险显著增加。因此,本文聚焦的核心问题为:如何设计既能抵抗拜占庭攻击,又能适应极度不平衡数据分布的恶意流量检测机制。
本文采用经典的拜占庭攻击模型[4]。假设$ N $个客户端中,至多$ m $个客户端被攻击者攻陷成为拜占庭节点,其中$ m<N/2 $以保证诚实客户端占多数。恶意客户端集合记为$ \mathcal{M} $,$ \left| \mathcal{M}\right| \leq m $;诚实客户端集合记为$ \mathcal{H} $,$ \left| \mathcal{H}\right| =N-m $。拜占庭攻击的数学形式化定义如下:对于每个恶意客户端$ h_i\in \mathcal{M} $,其上传至服务器的梯度 $ \tilde{\boldsymbol g}_{i}^{t} $不一定等于真实的本地梯度$ \boldsymbol g_{i}^{t}=\nabla {F}_{i}\left({\boldsymbol w}^{t}\right) $。攻击者根据攻击策略$ \mathcal{A} $构造恶意梯度。本文主要关注以下8类攻击。
1)符号翻转攻击(SignFlip[24],SF):$ {g}_{i}=-c\cdot \nabla {F}_{i}\left(\theta \right) $,其中$ c>0 $为放大系数,攻击将本地梯度乘以负常数以实现方向完全翻转。
2)部分翻转攻击(Partial,PT):仅对部分参数坐标进行符号翻转,使恶意梯度集合内部方向混杂。
3)高斯噪声攻击(Gaussian[24],GS):$ {g}_{i}=\nabla {F}_{i}\left(\theta \right)+\epsilon $,其中$ \epsilon \sim \mathcal{N}(0,{\sigma }^{2}I) $为加性高斯噪声。
4)范数放大攻击(Boosting,BT):$ {g}_{i}=c\cdot \nabla {F}_{i}\left(\theta \right) $,其中 $ {c} $ 为极大放大系数,使恶意梯度主导聚合方向。
5)定向攻击(TrimAware):$ {g}_{i}=c\cdot 1 $,刻意制造极端梯度值以拉高均值,针对TrimMean优化设计而提出。
6)统计模仿攻击(LIE[25]):$ \tilde{\boldsymbol g}_{i}^{t}={\mu }^{t}+z\cdot {\sigma }^{t} $,其中$ {\mu }^{t} $和$ {\sigma }^{t} $分别为所有诚实客户端梯度的均值和标准差,$ z>0 $为微调系数。攻击将每个坐标设置为均值向正方向偏移$ z $个标准差,使恶意梯度在统计上接近诚实梯度。
7)约束优化攻击(Min-Max[26],MM):通过约束优化最大化攻击效用的同时逃避检测,$ \tilde{\boldsymbol g}_{i}^{t}= \mathop{{\rm arg}\;{\max}}\limits_{g}\| {w}^{t+1}\left(\boldsymbol g\right)-\boldsymbol w_{\rm honest}^{t+1}\| $,约束条件为恶意梯度落在诚实梯度凸包内,即$ \boldsymbol g\in Conv\left(\{\boldsymbol g_{j}^{t}{\}}_{j\in \mathcal{H}}\right) $。
8)自适应攻击(Adaptive,AD):本文的自适应攻击策略为根据历史全局模型动态调整扰动策略,以KL散度度量跨轮次梯度行为偏差。该攻击为逃避基于单轮梯度统计特性的检测,刻意使恶意梯度在方向上与良性梯度保持近似一致,但其跨轮次扰动策略的频繁切换导致恶意梯度内部的离散程度显著高于良性梯度,此特征为梯度指纹辨识提供了可观测依据。
上述攻击对梯度统计特性的扰动模式各异:SF改变梯度方向,BT放大梯度范数,LIE和MM操纵梯度分布的偏度和峰度。这些特性决定了其防御方法,也成为其攻击类别的识别手段。
本文以8种代表性防御聚合方法作为基线。
1)FedAvg[2]。标准加权平均聚合,$ {\boldsymbol w}^{t+1}=\displaystyle\sum\limits_{i\in {\mathcal{S}}^{t}}\dfrac{{n}_{i}}{\displaystyle\sum\limits_{j}{n}_{j}}\cdot $$ \boldsymbol w_{i}^{t+1} $。无拜占庭鲁棒性,作为无防御参照基线,代表无防护条件下的性能下界。
2)Krum[8]。在$ S $个候选更新中,选择与其他更新具有最小平方欧氏距离和的梯度。形式化地,对每个梯度$ {g}_{i} $,定义得分$ s\left(i\right)=\displaystyle\sum\limits_{i'\in \mathcal{S}\left(i;S-m-2\right)}\| {g}_{i}-{g}_{i'}{\| }^{2} $,其中$ \mathcal{S}\left(i;k\right) $为与$ {g}_{i} $距离最近的$ k $个梯度的索引集合。Krum选择$ {{g}}_{{{{i}}^{*}}} $ 使得$ {i}^{*}=\mathop{\arg{\min}}\limits_{i}s\left(i\right) $,全局更新为$ {w}^{t+1}={w}^{t}-\eta \cdot {g}_{{{i}^{*}}} $。
3)Median[8]。在每个坐标维度$ j\in \{1,… ,d\} $上独立计算所有候选梯度的中位数:$ w_{j}^{t+1}=median\left(\{w_{i,j}^{t+1}{\}}_{i\in {{\mathcal{S}}^{t}}}\right) $。该方法对极端异常值具有天然的鲁棒性。
4)TrimMean[9]。在每个坐标维度上,剔除该坐标上取值最大和最小的$ \lfloor c\cdot S\rfloor $个值后计算均值,其中$ c\in \left(0{,}1/2\right) $为修剪比例。
5)NormClip[10]。首先计算所有候选梯度的$ {L}_{2} $范数中位数$ m=median\left(\{\| {g}_{i}{\| }_{2}\}\right) $,然后对每个梯度执行范数裁剪:$ {\hat{g}}_{i}={g}_{i}\cdot \min\left(1,\tau \cdot m/\| {g}_{i}{\| }_{2}\right) $,其中$ \tau \gt 0 $为裁剪阈值系数,最后取裁剪后梯度的均值作为全局更新。
6)FedProx[11]。在本地目标函数中添加近端正则项:$ F_{i}^{\rm prox}\left(w\right)={F}_{i}\left(w\right)+\dfrac{\mu }{2}\| w-{w}^{\rm t}{\| }^{2} $,通过限制本地更新与全局模型的偏离程度来降低恶意更新的影响。
7)Multi-Krum[8]。Krum的扩展版本,在每轮中选择多个可信梯度进行聚合,而非仅选择单个最优梯度,在提高鲁棒性的同时降低方差。
8)Bulyan[13]。首先通过Multi-Krum筛选$ N-2m $个可靠梯度,再执行坐标级TrimMean裁剪,理论上兼具选择鲁棒性与裁剪稳定性。
上述防御方法的鲁棒性边界各不相同:Krum在$ m< S/2 $时具有理论收敛保证;Median和TrimMean假设正常梯度在坐标级上集中分布;NormClip假设恶意梯度的范数显著大于正常梯度;FedProx通过近端约束间接限制恶意更新的幅度。本文提出的GF-FedIDS框架利用上述典型攻击的梯度特征扰动模式,建立攻击类型到最优防御方法的自适应映射。
考虑到不同拜占庭攻击对梯度特征施加的扰动模式具有显著差异:SF攻击反转梯度方向,BT攻击放大梯度范数,LIE攻击模仿良性统计特性,而自适应攻击则动态调整扰动策略。同时,不同防御方法对特定扰动模式的抵抗能力亦各不相同:中位数对方向反转免疫,范数裁剪对异常范数有效,Krum对统计孤立点敏感。本文提出GF-FedIDS框架,不再依赖通用防御对梯度的无差别过滤,而是通过精确识别攻击类型,定向选择只针对当前攻击有效的防御方法,从而在消除拜占庭扰动的同时最大程度保留攻击类别的梯度信息。为此,本文通过解析恶意梯度与良性梯度在方向、范数和分布方差三维特征上的数学关系构建“梯度指纹”,以作出最优拜占庭攻击防御选择,并结合类别感知加权机制,以感知处理动态网络中海量异质终端入侵检测数据的不平衡性,提升整体性能。整体框架如图1所示。
本节提出梯度指纹,并从理论上刻画每种攻击对梯度特征的扰动模式,并推导对应的防御最优性。
定义1. 梯度指纹。设第$ {t} $轮通信中,良性客户端集合为$ \mathcal{B} $,恶意客户端集合为$ \mathcal{M} $。定义良性梯度均值$ {\boldsymbol g}^{\rm ben}=1/\left| \mathcal{B}\right| \displaystyle\sum\limits_{i\in \mathcal{B}}{\boldsymbol g}_{i} $与恶意梯度均值$ {\boldsymbol g}^{\rm mal}=1/\left| \mathcal{M}\right| \displaystyle\sum\limits_{i\in \mathcal{M}}{\boldsymbol g}_{i} $。梯度指纹作为一个三维空间内的点可以被定义为:$ f=\left(CS,NR,VR\right) $。
其中,余弦相似度$ CS={\boldsymbol g}^{\rm mal}\cdot {\boldsymbol g}^{\rm ben}/\| {\boldsymbol g}^{\rm mal}\| \cdot \| {\boldsymbol g}^{\rm ben}\| $,度量恶意梯度与良性梯度的方向一致性;范数比$ NR= \| {\boldsymbol g}^{\rm mal}\| /\| {\boldsymbol g}^{\rm ben}\| $,度量恶意梯度的相对强度;方差比$ VR= Var\left(\{{\boldsymbol g}_{i}\colon i\in \mathcal{M}\}\right)/Var\left(\{{\boldsymbol g}_{i}\colon i\in \mathcal{B}\}\right) $,度量恶意梯度内部离散程度相对于良性梯度的比值。
基于上述梯度指纹特征,本文建立如表1所示的攻击类型与最优防御的映射。
相对各类攻击类型的最优防御理论论证如下。
1)SF攻击$ \rightarrow $Median。SF攻击将本地梯度乘以负常数$ -c $($ c>0 $),即$ g'=-c\cdot g $。对于中位数聚合,考虑各维$ j $的梯度值集合 $ \{{g}_{1,j},{g}_{2,j},… ,{g}_{n,j}\} $(含$ m $个恶意值$ \{-c\cdot {g}_{i,j}\} $)。当$ c> 1 $时,恶意值$ \{-c\cdot {g}_{i,j}\} $位于数轴上与良性值相反方向且距离更远的位置。设良性梯度的中位数为$ {med}_{\rm ben} $,由于恶意值偏离中心区域,它们不会翻转中位数的符号。形式化地,对于足够大的良性集合,$ med\left(\{-c\cdot {g}^{\rm mal}\}\cup \{{g}^{\rm ben}\}\right)=med\left(\{{g}^{\rm ben}\}\right) $,即中位数对反转操作免疫。
2)PT攻击$ \rightarrow $Median。PT攻击仅对部分坐标进行符号翻转,导致恶意梯度集合内部方向混杂。设翻转比例为$ p\in \left(0{,}1\right) $,则恶意梯度可表示为$ g'= (1- 2p )\cdot g+\delta $,其中$ \delta $为翻转噪声。此时$ g' $的方向介于$ g $和$ -g $之间,CS$ \in \left[-0.3{,}0.1\right] $。中位数在各维度独立操作,不受方向混杂的影响,其崩溃点(breakdown point)为$ {\epsilon }^{*}=0.5 $,在恶意比例低于50%时保持稳健。
3)GS攻击$ \rightarrow $NormClip。GS攻击向梯度注入零均值高斯噪声:$ g'=g+\xi $,$ \xi \sim \mathcal{N}(0,{\sigma }^{2}I) $。其期望范数为$ \mathbb{E}\left[\| g'{\| }^{2}\right]=\| g{\| }^{2}+d{\sigma }^{2} $,其中$ d $为参数维度。当$ d $较大时(网络流量特征的高维性),$ \| g'\| $以高概率显著偏离良性范数。范数裁剪操作 $ {\mathit{\Pi }}_{\tau }\left(g'\right)=g'\cdot \min (1, \tau /\| g'\| ) $严格限制$ \| {\mathit{\Pi }}_{\tau }\left(g'\right)\| \leq \tau $,将噪声引起的范数膨胀压制在阈值以内。
4)BT攻击$ \rightarrow $TrimMean。BT攻击将良性梯度放大$ {c} $倍:$ g'=c\cdot g $($ c\gg 1 $),使得恶意梯度在数值上成为极端大值。$ \beta $-截断均值首先剔除排序后最大的$ \beta n $个和最小的$ \beta n $个梯度,再对剩余梯度取平均。当恶意客户端比例小于等于$ \beta $时,所有放大梯度被剔除,聚合结果等于良性梯度的均值,攻击效果归零。
5)TrimAware攻击$ \rightarrow $TrimMean。TrimAware攻击刻意制造极端梯度值以拉高均值,其设计针对传统Mean聚合。由于攻击者产生的是定向极端值(全部偏大或偏小),恰好落入截断均值的剔除区域。设攻击产生$ m $个值为$ {g}_{\max} $的梯度,当$ m\leq \beta n $时,$ \beta $-TrimMean完全过滤这些极端值。
6)LIE攻击$ \rightarrow $Krum。LIE攻击使恶意梯度模仿良性梯度的统计特性,即$ g'={\mu }_{\rm ben}+{\sigma }_{\rm ben}\cdot \xi $。虽然单点统计特性相似,但LIE攻击者无法同时与所有良性梯度保持高相似性。Krum算法[6]计算每个梯度与其余$ n-f-2 $个最近邻距离之和,选择最小者。设良性梯度集合$ \{{g}_{i}{\}}_{i\in \mathcal{B}} $的内部距离为$ {d}_{in} $,恶意梯度$ g' $到各良性梯度的距离为$ {d}_{i}=\| g'-{g}_{i}\| $。由于$ g' $为单一统计模仿点,$ \displaystyle\sum\limits_{i\in S}{d}_{i} $($ \left| S\right| =n-f-2 $)必然大于某良性梯度与其邻居的距离之和,故Krum可选择良性梯度。
7)MM攻击$ \rightarrow $Krum。MM攻击通过约束优化最大化攻击效用的同时逃避检测。设约束为 $ \| g'- {g}^{\rm ben}\| \leq \delta $,则$ g' $位于以$ {g}^{\rm ben} $为中心、$ \delta $为半径的球内。此类优化后的梯度在高维空间与其他良性梯度保持可控距离$ \mathit{\Theta }\left(\delta \right) $,而良性梯度之间距离为$ \mathit{\Theta }\left({\sigma }_{\rm intra}\right) $。当$ \delta $与$ {\sigma }_{\rm intra} $量级相当时,MM梯度在Krum评分中表现为孤立点,因其无法同时与足够多的良性梯度形成紧密邻域。
8)AD攻击$ \rightarrow $组合防御。自适应攻击根据历史全局模型动态调整扰动策略,其梯度特征随轮次变化。本文通过梯度指纹的动态变化模式($ VR\gg 1 $且CS波动)识别自适应攻击,采用NormClip限制范数波动叠加TrimMean剔除极端值的组合策略。
定理1. 梯度指纹可分性。对于表1所列8种标准拜占庭攻击,设其三维梯度特征向量分别为 $ {\boldsymbol f}_{1}, {\boldsymbol f}_{2},… ,{\boldsymbol f}_{8}\in {\mathbb{R}}^{3} $。在良性梯度满足亚高斯分布$ {g}^{\rm ben}\sim subG ({\sigma }^{2} ) $且恶意客户端比例$ \beta <0.5 $的条件下,存在线性超平面$ {H}_{ij}\colon a_{ij}^{\rm T}\boldsymbol f+{b}_{ij}=0 $使得对任意$ i\neq j $,$ {\boldsymbol f}_{i} $和$ {\boldsymbol f}_{j} $分别位于$ {H}_{ij} $两侧,即梯度指纹在三维特征空间中线性可分。
证明. 由表1可知,8种攻击的CS覆盖范围互不重叠:SF/PT为负值区间$ \left[-1{,}0.1\right] $,GS近似为0$ \left[-0.1{,}0.1\right] $,LIE/MM为中等正值($ \left[0.2{,}0.6\right] $),BT/TrimAware为较高正值($ \left[0.3{,}0.9\right] $)(其中TrimAware的NR显著高于BT),AD在0附近但VR显著高于其他攻击。NR维度进一步区分GS(高NR)、LIE(低NR)和MM(低NR)。进一步结合VR维度,各类攻击的特征向量位于$ {\mathbb{R}}^{3} $的不同区域,由超平面分离定理即得线性可分性,详细证明见附录A。 证毕.
定理2. 防御最优性。设攻击类型为$ \mathcal{A} $,对应的梯度指纹为$ {f}_{\mathcal{A}} $,通过表1映射选择的防御为$ {d}^{*}= \phi \left({f}_{\mathcal{A}}\right) $。定义聚合后梯度偏差为$ \mathit{\Delta }\left(\boldsymbol w,\boldsymbol D\right)=\| {g}^{\rm agg}\left(\boldsymbol D\right)- {g}^{\rm ben}\| $,则$ {\boldsymbol D}^{*} $满足:$ \mathit{\Delta }\left(\boldsymbol w,{\boldsymbol D}^{*}\right)\leq \mathit{\Delta }\left(\boldsymbol w,\boldsymbol D\right) $,$ \forall {\boldsymbol D}\in \mathcal{D}=\{\rm Median, \rm TrimMean,Krum,NormClip\} $。
证明. 对每种攻击类型,所选防御$ {{\boldsymbol D}}^{*} $恰好消除了该攻击的扰动模式:Median免疫方向反转,NormClip限制噪声范数膨胀,TrimMean剔除极端值,Krum利用统计模仿的孤立性。其他防御对该攻击的特定扰动模式不具备同等消除能力,故偏差更大。详细证明见附录B。 证毕.
基于上述理论分析,本节给出GF-FedIDS框架中基于梯度指纹实现的决策算法描述,包含3个核心算法模块。
算法1. 梯度指纹提取算法。
输入:第$ {t} $轮所有本地客户端梯度集合$ \{{g}_{{i}}^{\left({t}\right)}\}_{{i}={1}}^{\boldsymbol{n}} $,恶意客户端索引集合$ \cal M $;
输出:各客户端梯度指纹$ \boldsymbol{f}= (CS,NR,VR ) $。
① for $ g_{i}^{\left(t\right)} \in \{g_{i}^{\left(t\right)}\}_{i=1}^{n} $ do
②  $ {g}^{\rm ben}\leftarrow \dfrac{1}{n-\left| \mathcal{M}\right| }\displaystyle\sum\limits_{i\notin \mathcal{M}}g_{i}^{\left(t\right)}; $/*计算良性梯度均值*/
③  $ {g}^{\rm mal}\leftarrow \dfrac{1}{\left| \mathcal{M}\right| }\displaystyle\sum\limits_{i\in \mathcal{M}}g_{i}^{\left(t\right)}; $/*计算恶意梯度均值*/
④  $ CS=\dfrac{{g}^{\rm mal}\cdot {g}^{\rm ben}}{\| {g}^{\rm mal}\| \cdot \| {g}^{\rm ben}\| }; $/*计算余弦相似度*/
⑤  $ NR=\dfrac{\| {g}^{\rm mal}\| }{\| {g}^{\rm ben}\| }; $/*计算范数比*/
⑥  $ {VR=}\dfrac{\dfrac{1}{\left| \mathcal{M}\right| }\displaystyle\sum\limits_{{i}\in \mathcal{M}}\| {{g}}_{{i}}-{{g}}^{\rm mal}{\| }^{2}}{\dfrac{1}{{n}-\left| \mathcal{M}\right| }\displaystyle\sum\limits_{{i}\notin \mathcal{M}}\| {{g}}_{{i}}-{{g}}^{\rm ben}{\| }^{2}}; $/*计算方差比*/
⑦ end for
⑧ return $ f=\left(CS,NR,VR\right) $。
算法2. 自适应防御选择算法。
输入:梯度指纹$ \boldsymbol{f}= ( CS,NR,VR ) $;
输出:最优防御方法$ {\boldsymbol{D}}^{*} $。
① 初始化$ {\boldsymbol{D}}^{*}\leftarrow \varnothing ; $
② if $ {CS} <{{\tau }}_{{CS}} $ then
③  $ {\boldsymbol{D}}^{*}\leftarrow $ Median$ ; $
④ end if
⑤ if $ {NR}>{{\tau }}_{{NR}} $ then
⑥  $ {\boldsymbol{D}}^{*}\leftarrow $ TrimMean;
⑦ end if
⑧ if $ {VR}>{{\tau }}_{{VR}} $ then
⑨ if $ {CS}\approx {0} $ then
⑩   $ {\boldsymbol{D}}^{*}\leftarrow $ {NormClip,TrimMean};
⑪ else
⑫   $ {\boldsymbol{D}}^{*}\leftarrow $NormClip;
⑬ end if
⑭ else $ {VR}<{{\tau }}_{{VR}} $ then
⑮ if $ {CS}>{{\tau }}_{{CS}} $ then
⑯   $ {\boldsymbol{D}}^{*}\leftarrow $ Krum;
⑰ else
⑱   $ {\boldsymbol{D}}^{*}\leftarrow $Median;
⑲ end if
⑳ end if
㉑ return $ {\boldsymbol{D}}^{*} $。
其中$ {\boldsymbol{D}}^{*}\in ${NormClip,TrimMean}表示执行NormClip进行范数裁剪以限制噪声幅度,再对裁剪后的梯度执行TrimMean聚合以剔除残留的极端值。该组合防御的具体执行方式为:给定当前轮所有梯度$ \{g_{t}^{(k)}\} $,首先按式$ \tilde{g}_{t}^{(k)}=g_{t}^{(k)}\cdot \min(1,\tau /||g_{t}^{(k)}||) $ 进行范数裁剪,再对裁剪后的梯度集合执行坐标级TrimMean聚合。上述决策树仅需3次浮点数比较即可完成防御选择,时间复杂度为$ {O}\left(1\right) $,计算开销可忽略不计。
梯度指纹决策树的阈值根据对基于攻击指纹的理论分布特性的离线分析确定,在本文中,CS阈值$ {\tau }_{CS} $取−0.3,NR阈值$ {\tau }_{NR} $取3.0,VR阈值$ {\tau }_{VR} $取1.5,简要的理论分析如下:
1)CS阈值$ {\tau }_{CS}。 $SF攻击将良性梯度符号反转,其理论$ CS\approx -1 $。设定方向反转的判定边界为$ {\tau }_{CS}= $$ -0.3 $,可有效区分方向性攻击与其他类型攻击,同时为高斯噪声等因素导致的轻微方向偏移保留适当容差。
2)NR阈值$ {\tau }_{NR} $。BT攻击将梯度放大$ \lambda $倍(典型值$ \lambda \in \left[5,20\right] $),产生$ NR\gg 1 $,$ {\tau }_{NR} $设定为3.0可识别范数显著偏离正常范围的攻击行为,同时避免将正常训练产生的较大梯度更新误判为攻击。
3)VR阈值$ {\tau }_{VR} $。TrimAware攻击通过放大梯度方差产生$ VR\in \left[1.5,3\right] $。方差异常的判定边界值$ {\tau }_{VR} $设定为1.5,可有效捕获此类精细化的方差操纵攻击。
上述阈值在实际部署中可根据本文思路酌情针对特定数据集或攻击类型特性进行调整,本文在设定上述阈值时实际执行的敏感性分析见附录C。
算法3. GF-FedIDS完整训练流程。
输入:本地训练数据集$ ({\boldsymbol{x}}_{{j}},{\boldsymbol{y}}_{{j}}) $,攻击样本比例$ {\alpha } $,恶意客户端比例$ {\beta } $,学习率$ {\eta } $,通信轮数$ {T} $;
输出:全局模型$ {\boldsymbol{w}}_{{T}} $。
① 初始化全局模型 $ {\boldsymbol{w}}_{{0}} $,计算类别权重${\boldsymbol{w}}_{\rm {attack}}\leftarrow $ $ \left({1}-{\alpha }\right)/{\alpha } $;
② for $ {t}\leftarrow {1} $ to $ {T} $ do
③ 广播$ {\boldsymbol{w}}^{{t-1}} $至所有客户端;/*本地训练*/
  /*本地训练阶段*/
④ for each $ {i}\leftarrow {1} $ to $ {n} $ 并行 do
⑤   $ \boldsymbol{w}_{{i}}^{\left({t}\right)}\leftarrow {\boldsymbol{w}}^{{t-1}}-{{\eta }}_{\rm {local}}\displaystyle\sum\limits_{{j}}{\boldsymbol{w}}_{{j}}\nabla l\left(\boldsymbol{f} ({\boldsymbol{x}}_{{j}};\boldsymbol{w} ),{\boldsymbol{y}}_{{j}}\right) $;
⑥   $ \boldsymbol{g}_{{i}}^{\left({t}\right)}\leftarrow {\boldsymbol{w}}^{{t-1}}-\boldsymbol{w}_{{i}}^{\left({t}\right)} $;/*计算本地梯度更新*/
⑦  上传$ \boldsymbol{g}_{{i}}^{\left({t}\right)} $至服务器;
⑧ end for
⑨ end for
  /*梯度指纹提取与策略选择阶段*/
⑩ if $ {\cal M}^{\left({t}\right)}\ne \varnothing $ then
⑪ 提取梯度指纹$ {\boldsymbol{f}}^{\left({t}\right)} $(算法1);
⑫ 选择最优防御$ {\boldsymbol{D}}^{*\left({t}\right)} $(算法2);
⑬ else $ {\boldsymbol{D}}^{*\left({t}\right)}{\leftarrow {{FedAvg}}} $;
⑭  $ \boldsymbol{g}_{{\rm agg}}^{\left({t}\right)}\leftarrow {\boldsymbol{D}}^{*\left({t}\right)}\left(\{\boldsymbol{g}_{{i}}^{\left({t}\right)}\}_{{i}={1}}^{\boldsymbol{n}}\right) $;
⑮  $ {\boldsymbol{w}}_{{t}}\leftarrow {\boldsymbol{w}}^{{t-1}}+{\eta }\cdot \boldsymbol{g}_{{\rm agg}}^{\left({t}\right)} $;/*更新全局模型*/
⑯ end if
⑰ return $ {\boldsymbol{w}}_{{t}} $。
动态网络中海量异质终端入侵检测数据的不平衡性[32](攻击样本比例$ \gamma \to 0 $)导致传统联邦学习优化目标严重偏向正常类。本节从理论上分析此问题,并提出类别感知实例加权(instance weighting,IW)机制。
设全局数据集中正常样本比例为$ 1-\gamma $,攻击样本比例为$ \gamma $,$ \gamma \to 0 $。传统联邦学习的优化目标为:
$ \begin{split} \underset{\theta }{\min}\,F\left(\theta \right)=&\left(1-\gamma \right)\cdot {\mathbb{E}}_{\left(x,y\right)\sim {{\mathcal{D}}_{0}}}\left[{ \ell}\left(f\left(x\right),y\right)\right]+\\&\gamma \cdot {\mathbb{E}}_{\left(x,y\right)\sim {{\mathcal{D}}_{1}}}\left[{ \ell}\left(f\left(x\right),y\right)\right] \;。\end{split} $
对$ \theta $求梯度:
$ \nabla F=\left(1-\gamma \right)\cdot \nabla {F}_{0}+\gamma \cdot \nabla {F}_{1} \;{,} $
当$ \gamma \to 0 $时,$ \nabla F\approx \nabla {F}_{0} $,攻击样本的梯度贡献趋于0,模型几乎不学习攻击类的判别特征,其中$\ell (\,) $为损失函数。
引入类别权重$ {w}_{1}=\left(1-\gamma \right)/\gamma $,构造新优化目标:
$ \begin{split} \underset{\theta }{\min}\,{F}_{\rm IW}\left(\theta \right)=&\left(1-\gamma \right)\cdot {\mathbb{E}}_{{{\mathcal{D}}_{0}}}\left[{ \ell}\right]+{w}_{1}\cdot \gamma \cdot {\mathbb{E}}_{{{\mathcal{D}}_{1}}}\left[{ \ell}\right] =\\&\left(1-\gamma \right)\cdot \left({\mathbb{E}}_{{{\mathcal{D}}_{0}}}\left[{ \ell}\right]+{\mathbb{E}}_{{{\mathcal{D}}_{1}}}\left[{ \ell}\right]\right) \;。\end{split}$
攻击样本对总梯度的贡献比例变为:
$ \begin{split} &\dfrac{{\gamma\cdot }{{w}}_{1}\cdot \| \nabla {{F}}_{1}\| }{\left({1-\gamma }\right)\cdot \| \nabla {{F}}_{0}{\| +\gamma \cdot }{{w}}_{1}\cdot \| \nabla {{F}}_{1}\| }=\\&\dfrac{\left({1-\gamma }\right)\cdot \| \nabla {{F}}_{1}\| }{\left({1-\gamma }\right)\cdot \| \nabla {{F}}_{0}\| +\left({1-\gamma }\right)\cdot \| \nabla {{F}}_{1}\| } =\dfrac{\| \nabla {{F}}_{1}\| }{\| \nabla {{F}}_{0}\| +\| \nabla {{F}}_{1}\| }\;。\end{split} $
此时2类梯度贡献的比例为$ \| \nabla {F}_{0}\| \colon \| \nabla {F}_{1}\| $,当2类梯度的期望范数同阶时,攻击样本对总梯度的贡献从$ {O}\left(\gamma \right) $提升至$ {O}\left(1\right) $,有效消除了类别不平衡导致的梯度湮灭效应。此外,IW机制使模型在ROC曲线上的操作点向高召回率区域移动,满足安全场景对漏报敏感的需求。
定理3. IW收敛性。在假设损失函数$ { \ell} $为$ \mu $-强凸且$ L $-光滑、随机梯度方差有界$ \mathbb{E}\left[\| \nabla {{ \ell}}_{i}-\nabla F{\| }^{2}\right]\leq {\sigma }^{2} $的条件下,GF-FedIDS算法引入类别权重后的全局模型收敛速度为:
$ \mathbb{E}\left[F\left({\theta }_{T}\right)\right]-{F}^{*}\leq O\left(\dfrac{L\sigma _{\rm IW}^{2}}{{\mu }^{2}T}+\dfrac{\kappa {G}^{2}}{\mu }\right) \;{,} $
其中$ \sigma _{\rm IW}^{2}=\max\{{\sigma }^{2},{\left(1-\gamma \right)}^{2}{\sigma }^{2}/{\gamma }^{2}\} $,$ \kappa $为条件数,$ G $为梯度上界。
证明. 引入类别权重后,IW仅改变各样本在损失函数中的系数,不改变损失函数的凸性和光滑性常数的阶。有效损失$ {{F}}_{\rm {IW}} $保持$ \mu $-强凸和$ L $-光滑($ {L}_{\rm IW}= $$ \max\{1,\left(1-\gamma \right)/\gamma \}\cdot L $)。应用标准联邦学习收敛分析框架,在步长$ \eta =O\left(1/L\right) $的设定下,经$ T $轮迭代后收敛界为上述形式。 证毕.
本节围绕4个核心研究问题,在3个公开网络安全数据集、8种拜占庭攻击类型和8种基线方法条件下,系统验证GF-FedIDS框架的有效性,系统验证GF-FedIDS框架在海量异质终端环境下的拜占庭鲁棒性。实验涵盖平衡与不平衡数据分布、分攻击类型防御效果、计算开销与消融实验等多维度评估,所有结果均为5次独立实验的平均值。
实验选用3个具有代表性的网络入侵检测数据集,数据集涵盖周期较久,能够反映的网络动态变化,覆盖传统网络与物联网大量设备,能够代表大量异质终端的网络行为。
1)UNSW-NB15数据集[33]。包含9类现代网络攻击的流量记录,实验从中抽取20000条样本,特征维度为42,涵盖流持续时间、协议类型、包长统计等流量特征。
2)CICIDS-2017数据集[34]。包含DDoS、端口扫描、Web攻击等14类网络入侵行为,实验抽取50000条样本,特征维度为78维,包含前向/后向包统计、流标志位等深度流量特征。
3)CIC-IoT-2023数据集[35]。面向物联网场景设计,涵盖新型IoT攻击类型,实验抽取50000条样本,特征维度为39。
实验在Python 3.10、PyTorch 2.0框架下进行。联邦学习系统模拟100个异质终端客户端,采用Dirichlet分布(α=0.5)进行Non-IID数据划分,以模拟真实场景下各终端数据分布的异质性。本地训练批次大小为32,学习率为0.01,优化器采用SGD(动量为0.9)。每轮通信随机选取10%客户端参与聚合,默认进行10轮联邦学习通信。攻击样本比例(AR)和恶意客户端比例(MCR)根据实验需求具体设置。
1)模型架构。本地客户端采用4层全连接神经网络作为轻量级网络入侵检测分类器,隐藏层维度分别为256、128、64,激活函数使用ReLU,输出层采用Softmax。输入层维度根据数据集特征数调整,输出层维度为2(正常/攻击二分类)。
2)评估指标。攻击检测率(attack detection rate,ADR)是指被正确识别的攻击样本占全部攻击样本的比例,直接反映模型对恶意流量的检出能力,是本实验的核心评估指标;准确率(ACC)是指正确分类样本占总样本的比例;误报率(FPR)是指正常流量被误判为攻击的比例。需要特别指出的是,在AR较低(如AR=5%)的极度不平衡场景下,ACC指标具有严重误导性(即使将所有样本判为正常也可获得95%以上ACC),故不平衡条件下的核心评估指标为ADR与F1-score。
本文围绕以下4个研究问题展开实验。
1)研究问题1:梯度指纹决策能否在单攻击场景下实现优于固定防御的检测率?
2)研究问题2:不同攻击样本比例(AR)条件下,GF-FedIDS的ADR表现如何?
3)研究问题3:在混合攻击场景下,GF-FedIDS能否保持鲁棒性?
4)研究问题4:GF-FedIDS各组件的有效性如何?计算开销是否具有优势?
为评估GF-FedIDS在面对单一类型拜占庭攻击时的防御效能,本文分别针对各类攻击应用每类防御方法进行实验。表2展示了在UNSW-NB15数据集上的各个攻击平均ADR(AR=10%, MCR=20%)。图2为各方法在3个数据集、8种攻击下的ADR热力图。
从表2和图2可以得出以下关键发现。
1)UNSW-NB15数据集上性能细粒度分析。从表2可见,基线方法在8种攻击下的平均ADR为0.413,而GF-FedIDS达到0.595,相较于基线最佳值(0.572)性能比为1.04。具体到各攻击类型:在SF攻击下,Median表现最优(0.637),GF-FedIDS达到0.645(提升1.3个百分点);在GS攻击下,NormClip表现最优(0.685),GF-FedIDS达到0.683(降低0.3个百分点);在LIE攻击下,Bulyan表现最优(0.621),GF-FedIDS达到0.644(提升3.7个百分点)。值得注意的是,GF-FedIDS在所有8种攻击上均达到或接近最优基线水平,没有出现显著劣化的情况。这是因为GF-FedIDS通过梯度指纹自动识别攻击类型并选择对应最优防御,相当于预先知道当前攻击类型并匹配最优策略,从而避免了固定防御在非匹配攻击上的性能损失。
从图2的ADR热力图可以直观地观察到:FedAvg、Median等单一策略在部分攻击上呈现深色低ADR区域(如Median对GS仅0.210,NormClip对SF仅0.230),而GF-FedIDS在所有攻击单元上保持均匀的浅色高ADR表现,验证了自适应策略选择的全谱覆盖能力。
2)跨数据集对比分析。图3展示了3个数据集上的ADR对比。CICIDS-2017数据集上所有方法的ADR普遍比UNSW-NB15高5~8个百分点(GF-FedIDS从0.595提升至0.621),这是因为CICIDS-2017具有78维特征,更高的特征维度为攻击检测提供了更丰富的判别信息,攻击信号在高位空间中更易区分。CIC-IoT-2023数据集上所有方法的ADR普遍比UNSW-NB15低3~5个百分点(GF-FedIDS降至0.561),这是因为IoT数据集的数据异构度更高,不同客户端间的数据分布差异加大了模型收敛难度,同时也增加了梯度指纹的噪声。重要的是,GF-FedIDS在3个数据集上相对于基线最佳值的平均性能比保持一致(分别为1.04,1.05,1.05),验证了方法的泛化能力。
为评估类别不平衡程度(以AR为指标)对各方法ADR的影响,以及GF-FedIDS的IW机制在不平衡条件下的补偿效果,本节在恶意客户端比例MCR=20%不变的条件下,改变AR比例为5%,10%,15%,20%,30%,50%,系统评估GF-FedIDS在不同不平衡程度下的ADR表现。表3给出了各AR条件下的ADR对比结果,图4展示了对应的ADR梯度曲线。
从表3和图3可以得出以下关键发现。
1)极端不平衡条件(AR=5%)分析。在AR=5%的极端条件下,UNSW-NB15数据集上基线平均ADR仅为0.178,基线最佳Krum为0.217,而GF-FedIDS达到0.448,性能比达2.06。这一显著提升主要来自IW机制的贡献:当AR=5%时,$ {{w}}_{\rm {attack}}=19 $,即攻击样本在损失函数中的权重是正常样本的19倍,有效迫使模型学习攻击类特征而非简单地将所有样本预测为正常。在CICIDS-2017数据集上,性能比提升更为显著(2.29倍),这是因为高维特征空间中攻击样本的梯度信号更强,IW机制的放大效应更明显。在CIC-IoT-2023数据集上性能比为2.10,虽略低于CICIDS-2017但仍远高于单攻击场景下的提升幅度。
2)AR=10%~20%性能比分析。在AR=10%时,3个数据集上的性能比分别为1.74、1.61和1.63。此时$ {{w}}_{\rm {attack}}=9 $,IW机制仍提供显著的类别平衡效果。随着AR增加至20%,性能比逐渐收窄至1.22~1.25。这是因为AR升高意味着攻击样本数量增多,类别不平衡程度降低,IW机制的边际增益递减。
3)AR=50%平衡条件分析。在AR=50%时,GF-FedIDS在3个数据集上的ADR均低于基线最佳值0.95~0.98倍。这一现象具有重要的理论意义:当类别完全平衡时,IW机制的权重系数为1,不再提供额外增益;而GF-FedIDS的梯度指纹提取和决策树推断引入了一定的估计误差(2%~5%的误判率),在极端平衡条件下这一误差无法被IW补偿,导致性能略低于理论上界。然而,在实际入侵检测场景中,AR=50%的条件几乎不可能出现,因此该结果不影响方法的实用价值。
IW机制在AR<30%的不平衡条件下提供显著增益,且AR越低增益越大,有效缓解了“高精度陷阱”问题。图2展示了ADR随AR变化的梯度曲线,GF-FedIDS在低AR区间保持平缓下降(从AR=5%的0.448降至AR=20%的0.494),而基线方法呈现陡峭下降趋势(从0.217降至0.404),直观验证了IW机制的稳定性优势。
为评估GF-FedIDS在面对混合攻击(多种攻击类型同时存在或自适应切换)时的防御能力,本节设计了4种混合攻击场景,具体构造方式如下。1)SF+LIE双攻击混合:将恶意客户端均分为2组,第1组执行SF攻击(梯度符号反转,$ {g}^{{\rm mal}}=-{g}^{\rm {ben}} $),第2组执行LIE攻击(基于良性梯度均值和方差构造统计伪装梯度),2组独立构造恶意梯度,良性客户端正常参与。2)GS+BT双攻击混合:第1组执行GS攻击(注入高斯噪声,$ {g}^{\rm {mal}}={g}^{\rm {ben}}+\mathcal{N} (0,{\sigma }^{2} ) $),第2组执行BT攻击(梯度放大,$ {g}^{\rm {mal}}=\lambda \cdot {g}^{\rm {ben}} $,$ \lambda \gg 1 $),模拟噪声注入与幅度放大并存的复杂威胁。3)8种随机混合:每轮为每个恶意客户端独立均匀随机分配一种攻击类型,模拟攻击者缺乏协同的混乱场景。4)4种自适应混合:攻击者每5轮评估1次当前攻击效果(以ADR为指标),若检测到ADR有下降趋势,则在后续轮次中切换为另一种攻击类型(从SF、LIE、TrimAware、BT中循环选取),模拟智能自适应对抗环境。上述设置能够有效反映出真实环境下的混合攻击场景。表4给出了混合攻击场景下的ADR对比结果,图5展示了可视化柱状图。
从表4和图5可以观察到下述特性。
1)混合攻击下的策略错配效应。从表5可以清晰观察到,所有固定防御策略在混合攻击场景下面临严重的“策略错配”问题。以UNSW-NB15数据集为例,在“SF+LIE”双攻击混合下,基线平均ADR仅为0.280,较单攻击场景下的基线平均(0.413)下降32%。这是因为针对SF优化的防御(如Median)对LIE攻击几乎无效,而针对LIE的防御(如Bulyan)对SF效果不佳。Multi-Krum和Bulyan作为集成方法,在混合攻击下也表现不佳(0.318和0.324),这是因为集成方法的子组件在面对混合攻击时产生内部冲突,无法形成一致的筛选决策。
2)GF-FedIDS的动态适应优势。GF-FedIDS在混合攻击场景下展现出显著优势:在UNSW-NB15上,4种混合攻击的平均ADR达到0.587,相比基线最佳(0.329)性能比为1.78。其中,在“8种随机混合”攻击下性能比最为显著(2.02),这是因为攻击类型越多,固定防御策略的错配概率越高,而GF-FedIDS通过每轮的梯度指纹提取和攻击识别,动态切换至当前主导攻击对应的最优防御,有效避免了错配。在“4种自适应混合”攻击下性能比为2.09,这表明即使攻击者根据防御策略动态调整攻击方式,指纹检测仍能追踪攻击变化并做出相应调整。
3)跨数据集一致性。3个数据集上的平均性能比高度接近(1.78, 1.77, 1.81),表明GF-FedIDS的混合攻击防御能力具有良好的泛化性。图5的柱状图直观展示了各方法在混合攻击下的ADR对比,GF-FedIDS的ADR显著高于所有基线方法,且在不同混合攻击类型间保持稳定的优势。
为验证GF-FedIDS各组件的独立贡献,本文设计消融实验比较3种配置在UNSW-NB15数据集上的性能(8种攻击平均,AR=10%, MCR=20%)。同时对比平时时间开销来验证GF-FedIDS的效率。
1)组件贡献分析。GF-FedIDS(无IW)的ADR为0.515,低于固定最优基线(0.538)约4.3个百分点。这一差距揭示了梯度指纹估计误差的影响:在无先验知识的情况下,梯度指纹对攻击类型的识别并非100%准确(实测识别准确率约88%),约12%的误判导致策略选择次优,造成ADR损失。然而,引入IW机制后,完整GF-FedIDS的ADR提升至0.595(提升8个百分点),不仅弥补了梯度指纹误判带来的损失,还额外带来了5.7个百分点增益。这是因为IW机制提升了攻击类样本的梯度质量,使得攻击类良性梯度与恶意梯度的区分度增强,间接提高了梯度指纹的判别准确率,形成了“IW改善梯度质量→GF识别更准确→防御更有效→模型收敛更好”的正反馈循环。
此外,从ACC和FPR指标看,GF-FedIDS+IW在提升ADR的同时保持了最高的准确率(0.87)和最低的误报率(0.028),说明IW机制并未引入过多的假阳性预测,而是通过精确的权重调整实现了类别间的平衡学习。
2)计算开销对比。 联邦学习场景中的终端设备(尤其是物联网设备)计算资源受限,因此聚合方法的计算开销是实际部署中的关键考量因素。表6给出了各方法的实际计算时间对比。
GF-FedIDS每轮训练耗时仅为0.52 s,与FedAvg(0.45 s)相比仅增加16%的计算开销,远低于Krum(0.82 s)、Multi-Krum(1.85 s)和Bulyan(2.20 s)。这一优异的实际表现源于以下2个因素:1)决策树推理的时间复杂度为$ {O}\left({n\log m}\right) $,其中$ {m}\leq 32 $为常数,实际耗时不足1 ms,可完全忽略;2)动态策略选择避免了不必要的多重计算,当识别当前攻击为GS/BT等幅度型攻击时,直接调用NormClip($ {O}\left({nd}\right) $复杂度),仅在必要时才调用Krum或TrimMean。相比之下,Multi-Krum和Bulyan无论当前攻击类型如何,均固定执行多轮Krum筛选,造成了大量冗余计算。
从可扩展性角度分析,当客户端数量n从10增至100时,GF-FedIDS的耗时增长趋势与Krum一致(均为$ O ({{n}}^{2} ) $项主导),但由于常数因子显著变小(无需迭代执行),实际耗时仍远低于集成方法。这使得GF-FedIDS适用于大规模边缘部署场景,而Multi-Krum和Bulyan在中等规模($ n>50 $)时已难以满足实时性要求。
本文提出了GF-FedIDS框架,构建了基于梯度三维指纹的攻击辨识理论,实现了从攻击类型到梯度特征再到最优防御的精准映射,并设计类别感知实例加权机制,形成面向真实入侵检测场景的自适应安全闭环。GF-FedIDS以梯度指纹引导防御策略路由选择,使极度隐蔽的攻击亦可被稳定捕获。针对网络流量中攻击样本极度稀疏的领域特性,GF-FedIDS引入类别感知加权机制通过代价敏感学习有效补偿了攻击类别的梯度贡献,在极端不平衡场景下将攻击检测率提升至基线方法的2.06~2.29倍。框架以决策树推理替代冗余聚合计算,以单位计算开销达成较高鲁棒,为资源受限的异质边缘联邦学习提供了可行方案。本研究深化了对抗条件下联邦学习网络入侵检测理论,为构建面向内生安全、可解释的海量异质终端网络入侵防御体系提供了可参考方法。

参考文献 引证文献
排序方式:
1
Xu Ke, Yi Qiu, Kun Yang. A tutorial on the Internet of things: From a heterogeneous network integration perspective[J]. IEEE Network, 2016, 30(2): 102−108
2
Brendan M, Eider M, Daniel R, et al. Communication-efficient learning of deep networks from decentralized data[C]//Proc of the 20th Int Conf on Artificial Intelligence and Statistics. Fort Lauderdale, New York: PMLR, 2017: 1273−1282
3
Amine F M, Othmane F, Leandros M, et al. Federated deep learning for cyber security in the Internet of things: Concepts, applications, and experimental analysis[J]. IEEE Access, 2021, 9: 138509−138542
4
Nguyen D C, Ding Ming, Pathirana P N, et al. Federated learning for Internet of things: A comprehensive survey[J]. IEEE Communications Surveys & Tutorials, 2021, 23(3): 1622−1658
5
de Oliveira J A, Gonçalves V P, Meneguette R I, et al. F-NIDS—A network intrusion detection system based on federated learning[J]. Computer Networks, 2023, 236: 110010
6
刘飚, 张方佼, 王文鑫, 等. 基于矩阵映射的拜占庭鲁棒联邦学习算法[J]. 计算机研究与发展, 2021, 58(11): 2416−2429
Liu Biao, Zhang Fangjiao, Wang Wenxin, et al. A Byzantine-robust federated learning algorithm based on matrix mapping[J]. Journal of Computer Research and Development, 2021, 58(11): 2416−2429 (in Chinese)
7
Xie C, Koyejo O, Gupta I. Fall of empires: Breaking Byzantine-tolerant SGD by inner product manipulation[C]//Proc of the 36th Conf on Uncertainty in Artificial Intelligence (UAI). Toronto, Canada: PMLR, 2020: 261−270
8
Blanchard P, El Mhamdi E M, Guerraoui R, et al. Machine learning with adversaries: Byzantine tolerant gradient descent[C]//Proc of the 31st Conf on Neural Information Processing Systems (NIPS). New York: Curran Associates, 2017: 119−129
9
Yin D, Chen Y, Kannan R, et al. Byzantine-robust distributed learning: Towards optimal statistical rates[C]//Proc of the 35th Int Conf on Machine Learning (ICML). Stockholm, Sweden: PMLR, 2018: 5650−5659
10
Muñoz-González L, Co K T, Lupu E C. Byzantine-robust federated machine learning through adaptive model averaging[DB/OL]. arXiv preprint, arXiv: 1909.05125, 2019
11
Li T, Sahu A K, Zaheer M, et al. Federated optimization in heterogeneous networks[C]//Proc of the 2nd Conf on Machine Learning and Systems (MLSys). Austin, TX: mlsys. org, 2020: 429−450
12
莫慧凌, 郑海峰, 高敏, 等. 基于联邦学习的多源异构数据融合算法[J]. 计算机研究与发展, 2022, 59(2): 478−487
Mo Huiling, Zheng Haifeng, Gao Min, et al. Multi-source heterogeneous data fusion based on federated learning[J]. Journal of Computer Research and Development, 2022, 59(2): 478−487 (in Chinese)
13
Guerraoui R, Rouault S. The hidden vulnerability of distributed learning in Byzantium[C]//Proc of the 35th Int Conf on Machine Learning (ICML). Stockholm, Sweden: PMLR, 2018: 3521−3530
14
Dritsas E, Trigka M. Federated learning for IoT: A survey of techniques, challenges, and applications[J]. Journal of Sensor and Actuator Networks, 2025, 14(1): 9
15
Xiang Hanlong, Zhang Xuyun, Xu Xiaolong, et al. Federated learning-based anomaly detection with isolation forest in the IoT-edge continuum[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2026, 22(1): 1−19
16
Zhang Zikai, Suman R, Xu Jiahao, et al. Federated learning for smart grid: A survey on applications and potential vulnerabilities[J]. ACM Transactions on Cyber-Physical Systems, 2026, 10(1): 1−26
17
Alhasawi Y, Alghamdi S. Federated learning for decentralized DDoS attack detection in IoT networks[J]. IEEE Access, 2024, 12: 42357−42368
18
Khan A N, Rizwan A, Ahmad R, et al. Hetero-FedIoT: A rule-based interworking architecture for heterogeneous federated IoT networks[J]. IEEE Internet of Things Journal, 2023, 11(4): 5920−5938
19
杨志刚, 王卓彤, 吴大鹏, 等. 物联网中带有隐私保护的鲁棒联邦学习研究[J]. 电子与信息学报, 2023, 45(12): 4235−4244
Yang Zhigang, Wang Zhuotong, Wu Dapeng, et al. Robust federated learning with privacy protection in Internet of things[J]. Journal of Electronics & Information Technology, 2023, 45(12): 4235−4244 (in Chinese)
20
肖雄, 唐卓, 肖斌, 等. 联邦学习的隐私保护与安全防御研究综述[J]. 计算机学报, 2023, 46(5): 1019−1044
Xiao Xiong, Tang Zhuo, Xiao Bin, et al. A survey on privacy protection and security defense of federated learning[J]. Chinese Journal of Computers, 2023, 46(5): 1019−1044 (in Chinese)
21
Zhao Yue, Li Meng, Lai Liangzhen, et al. Federated learning with non-iid data[J]. arXiv preprint, arXiv: 1806.00582, 2018
22
Kairouz P, McMahan H B. Advances and open problems in federated learning[J]. Foundations and Trends in Machine Learning, 2021, 14(1-2): 1−210
23
穆旭彤, 程珂, 宋安霄, 等. 抗拜占庭攻击的隐私保护联邦学习[J]. 计算机学报, 2024, 47(4): 842−861
Mu Xutong, Cheng Ke, Song Anxiao, et al. Privacy-preserving federated learning against Byzantine attacks[J]. Chinese Journal of Computers, 2024, 47(4): 842−861 (in Chinese)
24
Chen Weilun, Zhang Zhaoxiang, Hu Xiaolin, et al. Boosting decision-based black-box adversarial attacks with random sign flip[C]//Proc of the 16th European Conf on Computer Vision (ECCV). Cham: Springer, 2020: 276−293
25
Baruch G, Baruch M, Goldberg Y. A little is enough: Circumventing defenses for distributed learning[C]//Proc of the 33rd Conf on Neural Information Processing Systems (NeurIPS). New York: Curran Associates, 2019: 8632−8642
26
Shejwalkar V, Houmansadr A. Manipulating the Byzantine: Optimizing model poisoning attacks and defenses for federated learning[C]//Proc of the 28th Annual Network and Distributed System Security Symp (NDSS). San Diego: Internet Society, 2021: 1−18
27
高鸿峰, 黄浩, 田有亮. 基于多方计算的安全拜占庭弹性联邦学习[J]. 通信学报, 2025, 46(2): 108−122
Gao Hongfeng, Huang Hao, Tian Youliang. Secure Byzantine-elastic federated learning based on multi-party computation[J]. Journal on Communications, 2025, 46(2): 108−122 (in Chinese)
28
Fang Minghong, Cao Xiaoyu, Jia Jinyuan, et al. Local model poisoning attacks to Byzantine-robust federated learning[C]//Proc of the 29th USENIX Security Symp (USENIX Security). Berkeley, CA: USENIX Association, 2020: 1605−1622
29
Allouah Y, Farhadkhani S, Guerraoui R, et al. Fixing by mixing: A recipe for optimal Byzantine ML under heterogeneity[C]//Proc of the 26th Int Conf on Artificial Intelligence and Statistics. New York: PMLR, 2023: 1232−1300
30
Zuo Shiyuan, Yan Xingrun, Fan Rongfei, et al. Byzantine-resilient federated learning employing normalized gradients on non-IID datasets[DB/OL]. arXiv preprint, arXiv: 2408.09539, 2024
31
Zhang Haitao, Yao Ziming, Zhang Luyang, et al. Denial-of-service or fine-grained control: Towards flexible model poisoning attacks on federated learning[J]. arXiv preprint, arXiv: 2304.10783, 2023
32
Thabtah F, Hammoud S, Kamalov F, et al. Data imbalance in classification: Experimental evaluation[J]. Information Sciences, 2020, 513: 429−441
33
Moustafa N, Slay J. UNSW-NB15: A comprehensive data set for network intrusion detection systems (UNSW-NB15 network data set)[C]//Proc of the 2015 Military Communications and Information Systems Conf (MilCIS). Piscataway, NJ: IEEE, 2015: 1−6
34
Sharafaldin I, Lashkari A H, Ghorbani A A. Toward generating a new intrusion detection dataset and intrusion traffic characterization[C]//Proc of the 4th Int Conf on Information Systems Security and Privacy (ICISSP). Funchal, Portugal: SciTePress, 2018: 108−116
35
Garcia S, Parmisano A, Erquiaga M J. IoT−23: A labeled dataset with malicious and benign IoT network traffic (Version 1.0. 0)[DS/OL]. Zenodo, 2020. http://doi.org.10.5281/zenodo.4743746
2026年第63卷第10期
85
40
0
引用本文
BibTeX
文章信息
doi: 10.7544/issn1000-1239.202660484
  • 接收时间:2026-06-06
  • 首发时间:2026-11-17
  • 出版时间:2026-10-01
补充材料
相关文章
文章信息
作者
出版历史
  • 收稿日期:2026-06-06
  • 修回日期:2026-07-02
基金
作者信息
    1国防科技大学计算机学院 长沙 410073
    2信息支援部队工程大学网络信息学院 武汉 430030
    3湖南大学网络空间安全学院 长沙 410082

通讯作者:

参考文献
分享链接
https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202660484
分享至
全文二维码

扫描看全文

引用本文
BibTeX
本文的引用情况
2种不同金属材料的力学参数
科
Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)
属
Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏