网络安全与数据治理2025，Vol.44Issue(4)：24-31,8.DOI:10.19358/j.issn.2097-1788.2025.04.004
联邦学习中基于NMSS和LoRA的鲁棒防御机制研究
Robust defense mechanisms in federated learning:a study based on NMSS and LoRA
伏欣国 1王龙 2刘丽泽 3王雷 4赵建坤1
作者信息
- 1. 中国电子信息产业集团有限公司第六研究所,北京 102209
- 2. 中国电子信息产业集团有限公司第六研究所,北京 102209||山东大学 网络空间安全学院,山东 青岛 266237
- 3. 西安电子科技大学 计算机科学与技术学院,陕西 西安 710071
- 4. 深圳市酷开网络科技股份有限公司,广东 深圳 518000
- 折叠
摘要
Abstract
This study addresses security threats in federated learning,including privacy leakage,data poisoning,and model tam-pering.A defense architecture that integrates Non-Malleable Secret Sharing(NMSS)and Low-Rank Adaptation(LoRA)is pro-posed.The scheme uses a three-server threshold verification mechanism and zero-knowledge proof technology to secure parameter shards during transmission and recovery.In addition,the method applies low-rank constraints and dynamic weighted aggregation to limit malicious interference and reduce communication overhead.Experiments on the CIFAR-10 and mini-ImageNet datasets verify that the method improves defense accuracy,reduces model error,and enhances system robustness.The results show that the scheme is practical and scalable for large-scale scenarios.The study concludes that the architecture offers an efficient and fea-sible technical solution for secure federated learning.关键词
联邦学习/隐私保护/投毒攻击/LoRAKey words
federated learning/privacy-preserving/poisoning attack/LoRA分类
信息技术与安全科学引用本文复制引用
伏欣国,王龙,刘丽泽,王雷,赵建坤..联邦学习中基于NMSS和LoRA的鲁棒防御机制研究[J].网络安全与数据治理,2025,44(4):24-31,8.
