| With the development of Internet technology becoming more and more mature,the network security problems caused by the explosive growth of the number of network devices and traffic data have become more and more serious,and we are facing significant challenges.Intrusion detection can monitor abnormal network behavior in real-time and block dangerous intrusion behavior in time,which is an important means to solve network security problems.However,intrusion detection based on traditional methods still has problems such as unsatisfactory detection performance,high communication costs,exposure of sensitive data during data transmission,and inability to effectively share data to form "data islands".Federated learning and deep learning can effectively solve the above problems.In the actual intrusion detection process,different industry and organizations generate heterogeneous data,and there is asynchrony in the transfer of model parameters,which lead to communication delays,which will reduce the performance of intrusion detection methods based on federated learning.To address the above problems,this thesis launched the following two aspects of research:1.Aiming at the problems of low detection rate,"data island",high communication cost,and privacy leakage in the current intrusion detection,this thesis proposes an intrusion detection model(CS-FL)based on federated learning and deep learning,where the local classifier is a hybrid model of a convolutional neural network and self-attention mechanism.The CS-FL model framework combines the advantages of low computation and high accuracy of CNN and the advantages of capturing global features and internal correlations of Self-Attention and trains the intrusion detection algorithm through the federated learning framework.Experimental results show that this model not only improves the detection rate of intrusion detection,but also reduces communication costs,eliminates "data islands",protects data privacy and security,and has certain application value in practical research.2.The method based on federated learning still has the problem of client data heterogeneity and federated communication delay in the actual intrusion detection process.To solve this problem,this thesis proposes an optimization based on the CS-FL intrusion detection framework.This thesis first proposes to add an elastic factor to the client to increase the personalized weight of the client,to maintain the characteristics of the client and optimize the CS-FL intrusion detection model.Secondly,the gradient delay migration is proposed to solve the gradient delay problem caused by communication delay.When the server aggregates the global model,for the client with delay,the delay gradient of the current round is calculated according to the gradient similarity algorithm and the gradient delay of the global model,and then the global aggregation is performed,thus optimizing the CS-FL intrusion detection model.After experimental comparison,it is proved that the performance of the optimized CS-FL model is better than that of the CS-FL model. |
