Research On Key Technologies Of Smart Grid Intrusion Detection For Federated Learning

The smart grid integrates advanced information and communication technologies,which can realize the automation,the intelligence and the high efficiency of the power system.At the same time,it can also ensure intelligent dispatching,intelligent distribution,intelligent control and intelligent supply of the power system,and enhance the reliability,the safety,and the economy of the power system.Meanwhile,with the rapid development of industrial informatization and intelligence,network security issues have become increasingly complex and variable,which imposes severe network threats to the smart grid.Traditional intrusion detection methods based on deep learning can effectively detect network attacks and have been well applied and developed,but the following problems have been exposed in the face of the complex and diverse network environments in smart grids.First,due to the lack of high-quality data samples and the imbalanced distribution of sample label categories,it results in low accuracy and high false alarm rate of the intrusion detection model;Second,the traditional solution transmits data to the cloud server for centralized training,which brings the problems of high communication overhead and data privacy disclosure.Furthermore,some power stations are often unwilling to share their original data.In view of these,this article aims to explore how to apply federated learning technology to solve these problems,and further optimize intrusion detection models based on federated learning to ensure the security of model parameters during iterative training,which can improve the detection performance of the model and accelerate the convergence speed of the model.In summary,the main research content of this article are as follows:(1)In order to solve the problems of insufficient high-quality data,imbalanced sample label distribution and privacy protection of power data in smart grid intrusion detection,a federated learning intrusion detection model for unbalanced data is proposed.Firstly,the scheme adopts the intrusion detection model framework based on federated learning,which indirectly extends the features of the data samples by sharing the model parameters instead of the raw data,and protects the security of the power flow data.Secondly,the CNN-LSTM neural network based on space-time features is used for intrusion detection model training,which can further improve the accuracy of model detection.Finally,the SMOTE algorithm is utilized to process the flow data,which can make the minority category samples distribution balanced.The experimental results show that the proposed scheme has good advantages of performance,compared with other schemes.(2)In order to further improve the training efficiency of the intrusion detection model based on federated learning,and prevent the privacy leakage of model parameters in the iterative training process,an intrusion detection model based on secure dynamic aggregation federated learning is proposed.Firstly,the model adopts a joint multi-party approach to construct a global intrusion detection model to improve the detection accuracy of the training model.Secondly,a secure communication protocol is designed to protect the transmission security of model parameters during the iterative training process.Finally,a similarity-based dynamic weighted aggregation federated learning algorithm is proposed,which dynamically selects well-performing local models based on similarity measurement and uses them as weights for weighted aggregation.Experimental results show that this scheme can not only accelerate the convergence of the federated intrusion detection model but also reduce a certain amount of communication overhead.