基于权重攻击的联邦学习防御方案  

Federated learning defense scheme based on weighted attack

作　　者：张志远 周晖[1] ZHANG Zhi-yuan;ZHOU Hui(School of Information Science and Technology,Nantong University,Nantong 226019,China)

机构地区：[1]南通大学信息科学技术学院,江苏南通226019

出　　处：《计算机工程与设计》2023年第12期3521-3528,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(05049291)。

摘　　要：为提高联邦学习中恶意模型检测的准确率和鲁棒性,提出一种基于权重攻击的联邦学习防御方案。基于局部离群因子算法设计异常检测模型,提出用于检测异常模型的异常评分;提出一种基于密度的异常检测算法计算每个局部模型的异常评分;利用异常评分在聚合过程中自适应调整每个局部模型的权值。仿真结果表明,所提方案检测恶意模型精准度有所提高,具有良好的收敛性和稳定性。To improve the accuracy and robustness of malicious model detection in federated learning,a federated learning defense scheme based on adaptive weighting was proposed.The anomaly detection model was designed based on the local outlier factor algorithm,and the anomaly score for anomaly detection model was innovatively proposed.A density-based anomaly detection algorithm was proposed to calculate the anomaly score of each local model.The weights of each local model were adjusted adaptively by using outliers.Simulation results show that the proposed scheme improves the accuracy of malicious model detection,and has good convergence and robustness.

关 键 词：联邦学习 模型聚合 局部离群异常因子 卷积神经网络 准确率 局部模型 信息安全 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]