重庆理工大学学报2025，Vol.39Issue(9)：139-148,10.DOI:10.3969/j.issn.1674-8425(z).2025.05.017
融合深度自动编码器的联邦学习恶意节点检测方案
Federated learning malicious node detection scheme integrating deep autoencoder
摘要
Abstract
Federated learning enables multiple client nodes to collaboratively train a global model while preserving data privacy.However,the central server cannot fully control the behavior of individual nodes,leaving the system vulnerable to malicious nodes uploading erroneous gradient updates and thus undermining the global model's performance.To address the issue,this paper proposes FedDA,a federated learning defense framework integrating deep autoencoders for malicious node detection.FedDA identifies malicious nodes by analyzing the gradient information from the local model's output layer and leverages deep autoencoders for feature extraction,data decoupling,and dimensionality reduction.A federated aggregation algorithm based on Mahalanobis distance is introduced to mitigate the impact of malicious updates.Experiments on MNIST and CIFAR-10 demonstrate FedDA outperforms defense methods like Mkrum,improving the defense success rate by up to 19.9%while maintaining global model accuracy comparable to FedAvg.FedDA effectively prevents malicious behaviors and maintains the training performance of the global model,making it suitable for diverse federated learning scenarios.s.关键词
联邦学习/深度自动编码器/马氏距离/检测方案/聚合算法Key words
federated learning/deep autoencoder/Mahalanobis distance/detection model/aggregation algorithm分类
信息技术与安全科学引用本文复制引用
张晓琴,曹泽宇,陆艳军,金西兴..融合深度自动编码器的联邦学习恶意节点检测方案[J].重庆理工大学学报,2025,39(9):139-148,10.基金项目
重庆市技术创新与应用发展专项重点项目(CSTB2022TIAD-KPX0054) （CSTB2022TIAD-KPX0054）
重庆理工大学研究生教育高质量发展项目(gzlcx20243154) （gzlcx20243154）
