摘要
基于目前的隐私增强联邦学习方法可能存在准确率下降与通信开销增加等问题,甚至可能产生新的不安全因素,提出了一种差分隐私增强的双鉴别器条件生成对抗网络模型。在该模型中,引入了双鉴别器结构,通过模型中生成器和不同鉴别器之间的两两博弈过程,使得生成器所生成的数据在满足差分隐私的要求的同时尽可能接近输入数据。在联邦学习框架中应用该模型,可以保证模型的准确率不会因为隐私保护措施而大幅下降,与此同时增强联邦学习隐私保护的能力。通过仿真实验验证了所提出方案在点对点架构下的有效性。
Given that the current privacy-enhanced federated learning methods can encounter issues such as reduced accuracy and increased communication overhead and may even introduce new safety risks,a differential privacy-enhanced dual discriminator conditional generative adversarial network model is proposed.In this model,a dual discriminator architecture is introduced.Through the pairwise game between the generator and different discriminators in the model,the data generated by the generator is as close to the input data as possible while meeting the requirements of differential privacy.When applied in the federal learning architecture,this model\ensures that the accuracy of the model is not significantly compromised by privacy protection measures while simultaneously enhancing the privacy protection capabilities of federal learning.Simulation results verify the effectiveness of the proposed scheme in Peer-to-Peer architecture.
作者
沈翰林
汪学明
SHEN Han-lin;WANG Xue-ming(College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)
出处
《计算机工程与设计》
北大核心
2025年第8期2226-2232,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61163049)
贵州省自然科学基金项目(黔科合J字[2014]7641)。
关键词
联邦学习
隐私增强
准确率
差分隐私
双鉴别器结构
条件生成对抗网络
点对点构架
federal learning
privacy enhancement
accuracy
differential privacy
dual discriminator architecture
conditional generative adversarial network
peer to peer architecture
