Research On Intrusion Detection Based On Federated Learning For Smart Grid AMI Network

Adavanced Metering Infrastructure(AMI)is a communication system for two-way exchange of data and information between the power system and users,which ensures the normal operation of the smart grid.AMI transmits power data,control command data and other information.The power data includes user number,smart meter number,current total power and other data information.If the power data is attacked,the user’s power consumption and living habits such private information will be leaked;and the control command will affect the measurement work of the smart meter after being attacked,resulting in data measurement errors,thereby affecting the normal operation of the entire AMI.Therefore,it is necessary to detect network attacks in smart grid AMI network in time and study its intrusion detection method.However,the existing AMI network intrusion detection is difficult to balance between network attack detection and user privacy.Federated learning can perform intrusion detection on the premise of protecting user privacy.Therefore,this paper studies the intrusion detection based on federated learning in smart grid AMI network.The work of this paper is as follows:(1)In view of the privacy leakage problem in the long-distance transmission and centralized processing of data in smart grid AMI network,this paper proposes an AMI network intrusion detection method based on federated learning client selection and deep belief network with attention mechanism.In this method,the federated learning clients use local data and the model issued by the data center for training,and cooperate with the data center to build a horizontal federated learning framework.Under the premise of protecting data security by keeping data on the client for training,we design a client selection algorithm based on factors such as client computing power,communication quality,security risks,etc,which can improve the efficiency of federated learning;then we deploy a deep belief neural network with attention mechanism in each client to accurately detect possible network attacks in AMI network in real time.Experimental results show that the method can not only maintain privacy,but also efficiently detect network attacks,reducing communication and time overhead.(2)In the intrusion detection method based on federated learning,the inference attack is easy to leak the client parameters in the intrusion detection,and the poisoning attack destroys the client detection effect.To solve this problem,this paper proposes an intrusion detection method based on the federated learning clients security.This method uses homomorphic encryption CKKS as the underlying technology to resist the AMI network reasoning attack based on federated learning.In addition,in order to resist the destruction of parameters by poisoning attacks,this paper calculates the similarity of the parameter vector direction between the local model trained by each client and the data center model on the basis of homomorphic encryption CKKS,and scales the calculated values as adaptive weights;then,the size of the local model parameter gradient is normalized to the same size as that of the data center model;finally,the adaptive weights and the normalized gradients are used for weighted average to form the global model parameters.The experimental results show that the proposed method can effectively resist inference attacks and poisoning attacks,and maintain good detection performance in AMI intrusion detection based on federated learning.