Research On Network Intrusion Detection Model Based On Parameter Optimization Algorithm

The new generation of information technology is gaining momentum and integrating with various industries,producing a number of new products,technologies and models.Network intrusion detection,as the main technology to protect network system security,has been concerned and studied by scholars from all walks of life.At present,applying deep learning to intrusion detection is a developing trend in the whole field of network security.However,the existing intrusion detection technologies based on deep learning still have some problems,such as unbalance of network categories,long modeling time,high false positive rate,and easy to fall into local optimal parameters.According to the above problems existing in intrusion detection,this thesis gives specific solutions,research work and conclusions are as follows.(1)In this thesis,SMOTE oversampling method was used to extend the SMOTE data of a few categories in different proportion,so that most categories of data and a few categories of data in the network intrusion detection data set can achieve relative balance,which effectively solved the problem of imbalance of network data categories existing in the current intrusion detection techniques.In the experiment,UNSW-NB15 data set,which simulates real network attack environment,is selected to explain the principle and meaning of each type of attack feature in the data set,and the attack types are summarized and analyzed.The data that cannot be recognized by the neural network model is processed by the methods of data numeralization,standardization and normalization,and the multi-classification balanced data set that can be quickly recognized by the neural network model is constructed to effectively improve the classification quality.(2)Based on the analysis of network intrusion detection data,this thesis proposes a GSDE detection method aiming at the problems of low accuracy and high false positive rate existing in the existing intrusion detection technologies.By introducing the universal gravity operator into the differential evolution algorithm,the position information between particles in the population is improved,the optimization ability of the algorithm is improved,and the overall convergence rate of the algorithm is accelerated.In order to prove the effectiveness of the GSDE algorithm proposed in this thesis,three classical test functions,namely Sphere,Griewank and Rastrigin functions,are used for detection.By comparing the proposed GSDE algorithm with the traditional differential evolution algorithm,the experimental results show that the results of the three test functions are close to the optimal values.(3)Based on the proposed GSDE algorithm,this thesis constructs a multi-classification intrusion detection GSDE-GRU model to overcome the shortcomings of existing intrusion detection technologies such as long modeling time and easy to fall into local optimal parameters.By using the GSDE algorithm to continuously optimize the parameters of GRU model iteratively,the learning rate parameters of GRU model are optimized,so that the objective function is easier to converge to the extreme point,and the parameters are easy to fall into the local optimal situation.The experimental results show that compared with the classical GRU model,the accuracy rate of GSDE-GRU model proposed in this thesis is increased by 2%,the recall rate by 3.3%,the F1 value by 2.7%,the modeling time is shortened by 322 seconds,and the improvement of the identification results of attack categories with a small number of samples is particularly significant.Therefore,it is verified that the GSDE-GRU model proposed in this thesis has strong feature extraction ability,which can improve the training efficiency and detection performance of the model.