| Aiming at the problem that the traditional intrusion detection model has poor detection performance in high-dimensional data and data unbalanced environment,the feature learning and generalization ability based on deep learning model and the oversampling algorithm in the unbalanced algorithm are used to solve the problem of uneven data distribution.In this paper,an intrusion detection model combining adaptive oversampling algorithm(ADASYN,Adaptive Synthetic Sampling Approach)and improved stacked noise reduction self-encoder(SDA,Stacked Denoise AutoEncoder)is proposed.Based on the original deep neural network structure,the model algorithm is improved.Enhanced performance.First,the data oversampling process is performed using the ADASYN algorithm to generate new samples in categories that are difficult to classify.Secondly,the activation function replacement is performed by using the ELU with better effect.Then,the Adam optimization algorithm is used to replace the SGD random gradient descent method and the Dropout regularization.The two are combined to form an improved model of five-layer structure through SDA deep learning,and the low-dimensional and high-robust integration features are extracted.Finally,intrusion detection is performed in the softmax classifier.The comparison experiment of NSL-KDD dataset shows that the average accuracy of the model is 1.39%,4.69% and 4.82%,respectively,compared with SDA,AE-DNN and MSVM models.The average detection rate is increased by 6.74% and 9.62%,respectively.11.93%,the average false positive rate decreased by 0.58%,0.77% and 0.93%,respectively,effectively solving the problem of poor detection performance in massive high-dimensional data environment.At the same time,compared with the comparison model,the detection rates for small sample data U2 R and R2 L reached 100% and 97.5%,respectively,and the detection rate increased in the data imbalance environment.The Thesis has 26 pictures,17 tables,and 61 references. |
PDF Full Text Request