摘要
在数据隐私问题日益突出的背景下,联邦学习作为一种新型分布式机器学习技术,通过本地化数据处理和仅共享模型参数,降低了隐私泄露的风险。尽管其设计初衷是为了保护用户隐私,联邦学习本身仍面临一系列安全和隐私方面的挑战。回顾了联邦学习在安全性和隐私保护方面的最新研究进展,涵盖了其架构、常见攻击类型及防御策略。阐述了联邦学习的基本框架,并分析了常见的安全挑战,如数据中毒、模型中毒和拜占庭攻击等。讨论了隐私保护机制及其算法实现,评估其在提升安全性和隐私性方面的有效性,还发现某些攻击与防御技术之间存在潜在的转化利用关系,为攻防迁移研究提供了新思路。展望了联邦学习未来的发展方向,指出隐私保护算法优化、构建适应性强的隐私框架,以及推动隐私保护技术与多样化实际应用场景的深度融合,将成为推动联邦学习落地应用的关键挑战与发展机遇。
In response to growing concerns over data privacy,federated learning has emerged as a novel distributed machine learning paradigm that mitigates privacy risks by enabling local data processing and sharing only model parameters.However,despite its privacy-aware design,federated learning still faces a range of security and privacy challenges.This paper reviews recent research on federated learning security and privacy,including system architecture,major attack types,and defense strategies.It outlines the core framework of federated learning and analyzes threats such as data poisoning,model poisoning,and Byzantine attacks.It further discusses privacy-preserving mechanisms and their algorithmic implementations,evaluating their effectiveness.The study also highlights the potential transformation between certain attacks and defenses,providing insights into attack-defense transferability.Finally,future directions are discussed,focusing on the optimization of privacy-preserving algorithms,development of adaptive frameworks,and integration of privacy techniques into practical applications as key challenges and opportunities.
作者
平源
张云航
吴文红
潘志豪
康雯婷
刘宇建
PING Yuan;ZHANG Yunhang;WU Wenghong;PAN Zhihao;KANG Wenting;LIU Yujian(School of Information Engineering,Xuchang University,Xuchang 461000;School of Information Engineering,North China University of Water Resources and Electric Power,Zhengzhou 450046;Henan Province Engineering Technology Research Center of Big Data Security and Applications,Xuchang 461000;School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin 132022)
出处
《重庆邮电大学学报(自然科学版)》
北大核心
2025年第6期940-953,共14页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金项目(62162009)
河南省科技攻关计划(242102211065)
河南省创新型科技人才队伍建设工程(CXTD2017099)
河南省研究生教学质量工程项目(YJS2024AL112,YJS2024JD38)
许昌学院科研创新团队(2022CXTD003)。
关键词
联邦学习
安全攻击
隐私保护
防御策略
攻防迁移
federated learning
security attacks
privacy protection
defense strategies
attack-defense transferring
