Adversarial Attacks And Defense Methods For The Data-driven Algorithms In Cyber-physical Power Systems

With the construction of low-carbon power system and the development of power cyber-physical system(CPS),the power system has been characterized by high uncertainty and strong nonlinearity.The complexity of traditional model-driven analysis and control methods has increased dramatically,which gradually cannot meet the accuracy and real-time requirements in the new environment of power CPS,and the data-driven methods have become the effective solutions.However,the datadriven methods have some internal security risks,such as the data security risks and the algorithm security risks.The large-scale application of data-driven methods may transfer these security risks to the control process of power CPS,which results in new security problems across cyber space and physical space.This problem has not attracted enough attention,and the related research is in its infancy.A typical attack method against the data-driven algorithms in power CPS is the adversarial attack,which can induce the data-driven algorithms output error results by adding small perturbations to the input data.Compared with traditional cyber-attack methods,the adversarial attack is more hidden and can cause more serious consequences.In order to defend against the adversarial attacks in power CPS,it is necessary to analyze the characteristics of power CPS and the attack methods,and formulate effective defense strategies.Therefore,this paper is focused on the adversarial attack and defense methods of data-driven control strategies in power CPS.The main contents and innovations are as follows.Firstly,the potential weak points are screened according to the structure and characteristics of power CPS,and the attack vectors of adversarial attack are generated based on these weak points.Then,considering the limited attack resources,an attack vector injection method based on the distributed power supplies in user side is proposed,and the attack shame is optimized from the perspective of the attacker.According to the mechanism of adversarial attacks,the abnormal data identification and filtering methods are proposed,and based on the generative adversarial network(GAN)and data-knowledge integrating method,the robustness of the data-driven control strategy is improved.Finally,the power CPS cosimulation platform is constructed to simulate and verify the attack and defense methods.The specific work of this paper is as follows.(1)The structure and characteristics of power CPS are analyzed and the potential weak points are listed.Considering the attack expectations and constraints,the weak points screening method is proposed.An attack vector generator is constructed and trained by GAN,which can generate the attack vector in real-time according to the current state of the power CPS.This attack vector generation method provides theoretical and data support for the attack vector injection method in subsequent.(2)According to the characteristics of the attack vector and distributed power supply in user side,the model of disturbance source for injecting attack vector is constructed.A parameter identification method of attacked nodes and an attack vector fitting method are proposed,which can induce the distributed power supplies to inject the required attack vectors into power CPS by tampering with the data in the measurement module of the distributed power supplies.Considering the limited attack resources and defense resources,an attack scheme optimization method based on two-player attack and defense game is proposed.The above adversarial attack methods provide attack information support for the research on vulnerability discovering and defense methods in subsequent.(3)From the three aspects of abnormal data identification,vulnerability discovering,and robustness improving,the defense against adversarial attack of data-driven control strategy is carried out.For the input data with hidden attack vector,an abnormal data identification method based on clustering is proposed,and an input data filter is constructed to repair the data with attack vector.The robustness discrimination index of data-driven control strategy under adversarial attack is proposed,which considering the impact of adversarial attack,and the vulnerability discovering method based on GAN is proposed,which can effectively improve the robustness.For the defect of poor interpretability of data-driven methods,a data-knowledge integrating method is proposed,which uses the knowledge-driven model to improve the robustness of the integrating model,so that it can still output correct results for the input data with attack vectors.(4)The cosimulation platform based on MATLAB and OPNET is proposed to simulate and verify the adversarial attack and defense methods.Considering the advantages and disadvantages of the existing synchronization methods,a state-caching based synchronization method is proposed to balance the simulation accuracy and calculation speed.In addition,to solve the problems of processing multiple concurrent events and node mapping in power CPS modeling,a network boundary module is proposed that distributes and aggregates interface packets in accordance with the number of devices.The attack and defense experiments in this cosimulation platform prove the effectiveness of the methods in this paper,and highlight the necessity of studying adversarial attack and defense methods in power CPS.