Deep Learning Optimization Method Based On Genetic Algorithm And Its Application

In recent years,deep learning has been widely used and has become one of the most popular research fields.Deep Neural Network(DNN)and Convolutional Neural Network(CNN),as their main network models,have excellent feature extraction capabilities and high-dimensional data processing capabilities.However,the performance of DNN and CNN in intrusion detection is easily affected by network structure parameters such as initial weights and thresholds.Currently,the selection of these parameters depends on manual experience.On the same problem,different selection methods will lead to network models.Large differences in learning performance.With the continuous complexity of DNN and CNN,the method of manually selecting model structural parameters has become less and less practical.Based on genetic algorithm,this paper studies the structure and parameter optimization methods of DNN and CNN,and applies the optimized DNN and CNN to intrusion detection.This paper first proposes a genetic algorithm-based DNN structure and parameter optimization method and applies it to intrusion detection.Using genetic algorithm to optimize the number of hidden layers of DNN,learning rate and training times,etc.,to improve the learning and generalization ability of DNN.Design reasonable mutation rules for the number of hidden layers,learning rate,and training times to improve the convergence rate,diversity of groups,and prediction accuracy in finding the optimal value.At the same time,DNN model frameworks with different hidden layers are constructed to verify that the proposed method also has certain optimization performance after the number of hidden layers continues to increase.Finally,the pre-processed KDDCUP99 data set is used for experimental verification.The results show that the accuracy and detection rate of this method have been improved to a certain extent,and the false alarm rate has been improved compared to several unoptimized DNN structures and classic machine learning algorithms.This paper then gives a genetic algorithm-based CNN structure and parameter optimization method and applies it to intrusion detection.Use the powerful global optimization ability of genetic algorithms to automatically select the optimal CNN model structure with the optimal initial weights,thresholds,network structure,optimizer,and number of neurons in the fully connected layer to expand the diversity of CNN structures and reduce training time To automate the CNN structure.At the same time,CNN basic frameworks with different convolutional layers and pooling layers are constructed to verify that the proposed method also has certain optimization performance when the number of convolutional layers and pooling layers increases.Finally,the KDDCUP99 data set is used for experimental verification.The experimental results prove that this method has a significant increase in accuracy and detection rate,and a significantly reduced false alarm rate compared with the classic machine learning algorithm.Compared with the unoptimized CNN model,the accuracy and detection rate have been improved,and the false alarm rate has also improved.