| In recent years,with the continuous expansion of network scale and the development of various network technologies and protocols,the technology and frequency of network intrusion are also on the rise,which makes the information security of users face a huge threat.Intrusion detection technology plays an important role in maintaining network security,which can monitor the intrusion behavior in the network environment and generate feedback.Although the mainstream intrusion detection methods based on deep learning have high detection accuracy for the learned attack types,when there are unknown attack types or insufficient training data,a large number of labeled data need to be recollected and relearned.In view of the above problems,this paper mainly completed the following research work:1.An intrusion detection method based on ICNN fine-tuning is proposed.By fine-tuning the source domain training model with a few labeled target samples,the accuracy of model detection is improved and the cost of training calculation is reduced.Finally,the Accuracy of the test on KDDTest+ was improved by at least 17.08%,and that on KDDTest-21 was improved by 30.02%.2.An intrusion detection method based on ICNN-MMD is proposed.Considering the lack of target samples or even the absence of labels,the difference between source domain data and target domain data is substituted into neural network optimization as a loss to reduce the distribution difference between source domain and target domain to achieve domain adaptation.First,the source domain model is obtained by training the source domain data.Then,the Maximum Mean Discrepancy(MMD)is used to measure the distance between the source domain and the target domain,and the unsupervised secondary training is conducted to achieve domain adaptation and share the weights in the network.The experimental results on the NSL-KDD dataset show that the Accuracy of KDDTest+ is improved by 1.5%,and the Accuracy of KDDTest-21 is improved by 2.96%.3.A source domain data independent SFDA intrusion detection method is proposed.This method only uses the trained source domain model and unlabeled target domain data,freezes the source domain model and copies the parameters to the multi branch target model.Calculate the self-entropy value of each target sample after the prediction of the target model,screen the reliable sample set according to the self-entropy,calculate the category similarity score,and further filter the unreliable pseudo labels using the mechanism based on confidence filtering to jointly optimize the loss of self-learning branches and source domain regularization branches.The experimental results show that the Accuracy,Recall and F1 values of KDDTest+ increase by 1.65%,3.84% and 2.35% respectively. |
