Cyber-attacks Identification Based On Small Sample Data In Cyber Physical Power System

The promotion of smart grid construction makes the communication,calculation and control ability of power system develop greatly.With the development of information system and physical system,the relationship between them is getting deeper and deeper,and the overall system has cyber physical system features.The strong coupling relationship of information system and physical system improves the monitoring and control ability of power grid operation,but it also causes the risk of network attack.At present,the method mainly uses machine learning to detect network attacks,but this kind of network attack detection based on data mining faces great challenges in the unbalanced environment of small samples.In the face of the above difficulties in identifying attacks by traditional machine learning methods,small sample data of power CPS should be considered for learning and improving the algorithm.In this paper,the network attack identification method of CPS small sample data is studied:(1)The improved generative adversarial network is used to learn the internal distribution of small sample data and generate small sample data offline.In order to improve the quality of the original generated samples,the JS distance is replaced by Wasserstein distance.Finally,the effectiveness of the generated samples is verified by dimensionality reduction visualization and comparison of identification results before and after sample enhancement.(2)In view of the long training time of the generative adversarial network,which is difficult to meet the needs of fast feedback and real-time response,a network attack identification method based on MAD-GAN in enhanced data mode is proposed.Firstly,the long-term and short-term memory network is used to study the time series characteristics of power data,and small sample data is used to train the model offline.Secondly,the anomaly detection score is redesigned to build a fast online attack identification model based on MADGAN.(3)Under the premise of a certain amount of data needed to feed the generative adversarial network,considering the real system,if there is only a small amount of data,how to quickly identify the small sample of unbalanced attack data.To solve this problem,an attack identification method based on meta learning and deep forest is proposed.Firstly,the N-way K-shot method is used to construct training samples.Secondly,the network attack online identification is carried out by improving the deep forest model of cascaded forest structure,which provides a fast support for CPS network attack identification.