抗中毒攻击的鲁棒隐私保护模型  

The Robust Privacy Protection Model Against Poisoning Attacks

作　　者：李楠 叶嘉宾[1,2] 蒋彦甫 孙祥 刘雪慧 LI Nan;YE Jiabin;JIANG Yanfu;SUN Xiang;LIU Xuehui(Nanjing Big Data Group Co.,Ltd.,Nanjing,China,210000;Nanjing Intelligent Computing Technology Development Co.,Ltd.,Nanjing,China,210000;Southeast University,Nanjing,China,210000;Jimei University,Xiamen,China,361021)

机构地区：[1]南京大数据集团有限公司,江苏南京210000 [2]南京智能计算科技发展有限公司,江苏南京210000 [3]东南大学,江苏南京210000 [4]集美大学,福建厦门361021

出　　处：《福建电脑》2025年第3期20-23,共4页Journal of Fujian Computer

摘　　要：联邦学习在隐私保护方面容易受到恶意梯度中毒攻击,现有的防御策略存在计算和通信成本高的问题。为解决这些问题,本文提出一种抗病毒攻击的鲁棒隐私保护模型。该模型通过内部审计机制评估加密梯度,利用高斯混合模型结合马氏距离进行稳健聚合,并采用加性同态加密技术确保数据安全。实验结果显示,本文模型能有效抵御攻击,并显著降低计算和通信的成本,较其他加密技术表现出更好的隐私和准确性保障。Federated learning is susceptible to malicious gradient poisoning attacks in terms of privacy protection,and existing defense strategies suffer from high computational and communication costs.To address these issues,this article proposes a robust privacy protection model against viral attacks.This model evaluates encryption gradients through internal audit mechanisms,uses Gaussian mixture models combined with Mahalanobis distance for robust aggregation,and employs additive homomorphic encryption technology to ensure data security.The experimental results show that the model proposed in this paper can effectively resist attacks and significantly reduce the cost of computation and communication,demonstrating better privacy and accuracy protection compared to other encryption techniques.

关 键 词：隐私保护 中毒攻击 鲁棒隐私保护模型 

分 类 号：S762[农业科学—森林保护学]