摘要
随着工业物联网安全需求日益增长,如何在保护数据隐私的同时高效检测入侵攻击成为关键问题.联邦学习作为一种分布式机器学习方案,能够有效地进行数据隐私保护.然而,由于工业物联网设备所面临的数据异构性问题,传统的联邦学习方法FedAvg在训练过程中全局模型无法充分适应所有客户端的异构数据分布,从而影响全局模型的性能.为了解决这一问题,提出一种基于历史软标签的联邦知识蒸馏方法(FedKD-HSL).该方法在每轮训练中,各个客户端上传本地模型输出的软标签至中央服务器的缓存空间,从而生成具有全局数据分布特征的教师模型历史软标签;在此基础上,在新的训练周期内将教师模型历史软标签下发至各客户端,用于指导本地学生模型进行知识蒸馏,以有效促进全局知识向本地模型传递,缓解数据异构带来的模型性能下降问题.在公开数据集CIC-IDS2017上的实验结果表明,FedKD-HSL方法在三种不同数据异构程度的二分类及多分类任务中性能均显著优于传统的FedAvg算法.
As security demands in the Industrial Internet of Things(IIoT)continue to rise,achieving efficient intrusion detec-tion while preserving data privacy has become a critical challenge.Federated learning(FL)offers a distributed machine-learning paradigm that inherently protects data privacy.However,due to the data heterogeneity in IIoT environments,the traditional feder-ated learning algorithm FedAvg struggles to enable the global model to fully adapt to the heterogeneous data distributions across clients during training,which in turn degrades the overall performance of the global model.To address this issue,this paper pro-posed a federated knowledge distillation method based on historical soft labels,termed Federated Knowledge Distillation with Historical Soft Labels(FedKD-HSL).In each training round,clients uploaded the soft labels generated by their local models to a cache space on the central server,where historical soft labels with global data distribution characteristics were constructed and accumulated as teacher model output.Based on this,in subsequent training cycles,the historical soft labels were distributed to clients to guide knowledge distillation of local student models,thereby effectively promoting the transfer of global knowledge to local models and alleviating performance degradation caused by data heterogeneity.Experimental results on the public CIC-IDS2017 dataset showed that the proposed FedKD-HSL method significantly outperformed the traditional FedAvg algorithm in both binary and multi-class classification tasks under three different degrees of data heterogeneity.
作者
薛宏宇
李侯亮
黄琴霞
谢盈
XUE Hongyu;LI Houliang;HUANG Qinxia;XIE Ying(School of Computer Science and Artificial Intelligence,Southwest Minzu University,Chengdu 610041,China)
出处
《西南民族大学学报(自然科学版)》
2026年第1期78-89,共12页
Journal of Southwest Minzu University(Natural Science Edition)
基金
西南民族大学中央高校基本科研业务费专项资金资助(ZYN2025005)。
关键词
工业物联网
入侵检测
联邦学习
知识蒸馏
数据异构
Industrial IoT
intrusion detection
federated learning
knowledge distillation
data heterogeneity
