基于半监督联邦学习的分布式电源调控安全威胁协同检测

Collaborative Detection for Security Threats to Distributed Generator Regulation Based on Semi-supervised Federated Learning

针对现有分布式电源调控系统安全威胁检测存在的检测精度不高、通信效率低等突出问题,以及由于数据手工标注的高成本、低效率等客观原因和模型自动伪标注不可信导致的数据利用不充分问题,提出了一种基于半监督联邦学习(SSFL)的安全威胁分布式协同检测方法,在云端和边缘设备之间进行协同训练,并通过未标记数据进行模型自我学习和优化,从而更好适应分布式电源调控系统的安全威胁环境。首先,采用改进Transformer模型有效捕获安全威胁。其次,考虑到分布式电源调控系统的数据具有跨设备、跨区域的特点,引入联邦学习确保本地数据隐私安全。然后,针对未标记数据问题,通过云边协同训练获得全局模型并进行伪标记,设计一致性正则化与信息熵正则的损失函数以保证伪标记的可信度。最后,设计动态加权聚合方法优化参数更新和模型训练。在密西西比大学电力系统数据集上进行仿真实验,实验结果表明,与FedAvg-FixMatch方法和FedMatch方法相比,检测准确率分别提升了8%和4%,且类别召回率和精确率均有提高,显著减少了18%~28%的通信开销,表明了所提方法在分布式电源调控系统安全威胁检测中的有效性和实用性。

A collaborative detection method for regulation security threats to distributed generator based on semi-supervised federated learning is proposed to address the prominent issues of low detection accuracy and low communication efficiency in existing distributed generator regulation systems,as well as the objective reasons for the high cost and low efficiency of manual data annotation and insufficient data utilization caused by untrustworthy model automatic pseudo annotation.The method involves collaborative training between cloud and edge devices,and self-learning and optimization of the model are realized through unlabeled data to better adapt to the security threat environment of distributed generator regulation systems.Firstly,an improved Transformer model is used to effectively capture security threats.Secondly,considering the cross-device and cross-regional characteristics of data in distributed generator regulation systems,federated learning is introduced to ensure the privacy and security of local data.To address the issue of unlabeled data,a global model is obtained through cloud-edge collaborative training for pseudo labeling,and a loss function for consistency regularization and information entropy regularization is designed to ensure the credibility of pseudo labeling.Finally,a dynamic weighted aggregation method is designed to optimize parameter updates and model training.Simulation experiments are conducted on the power system dataset at the University of Mississippi,and the experimental results show that compared with FedAvg-FixMatch and FedMatch methods,the proposed method improved the detection accuracy by 8%and 4%,respectively,and both category recall and accuracy are improved,significantly reducing communication overhead by 18%~28%.This demonstrates the effectiveness and practicality of the proposed method in security threat detection in distributed generator regulation systems.