Research On Intrusion Detection Model Based On SMOTETomek And BiGAN

Generative adversarial network GAN is a common deep learning model,which can perfectly avoid the learning mechanism of Markov chain,and even has certain applicability when the probability density is not computable.Bidirectional Generative Adversarial Networks(BiGAN)are an improvement of GANs that enable GANs to learn important representations by adding encoders.In this thesis,an intrusion detection scheme based on Bidirectional Generative Adversarial Network(BiGAN)is proposed.This scheme can speed up data processing,improve detection accuracy,and solve data problems by optimizing BiGAN and SMOTETomek balancing processing of NSL-KDD dataset.Set overfitting and imbalance problems,thereby improving model performance.The main work is as follows:(1)Based on the NSL-KDD data set,the performance indicators such as running time and detection accuracy of BiGAN,AE,OCSVM,VAE,seq2seq,PCA and other models were experimentally analyzed and compared,and it was concluded that BiGAN is in these models.The running speed is relatively fast,and it can achieve an accuracy of 95.03%when the running time is 23.23s,which has a certain speed advantage.Experiments have proved that the experimental method in this p can maintain a high accuracy rate and also has a fast running speed.Therefore,the model based on this algorithm is selected as the research object of this thesis.(2)In order to improve the detection accuracy of the BiGAN model,the BiGAN model is optimized,and an optimized model SgBiGAN is given.The model uses softsign as the activation function to optimize the generator,which is further refined using the Adam optimization algorithm.(3)Aiming at the problem of class imbalance in the NSL-KDD dataset,resulting in low classification accuracy,the SMOTETomek combined sampling algorithm is used to balance the dataset.The experimental results show that using the processed data set for training and testing can not only improve the detection ability of the traditional model for the minority class of the sample under the effect of binary classification,but also increase the robustness of the model.(4)Further comparison experiments were conducted on the performance of the optimized model SgBiGAN and AE,OCSVM,VAE,Seq2seq,PCA and other models on the original data set and the data set processed by SMOTETomek,and various evaluation indicators were analyzed.The results show that the new model has better and stable performance on the balanced dataset,not only has high detection accuracy,but also can be created and run quickly.