At present,cybersecurity attacks have become an increasingly serious global problem,and the numerous cybersecurity incidents pose a major threat to social and economic development.In order to protect organizations and individuals from network attacks,intrusion detection techniques play an important role in ensuring information security,and the accurate identification of various attacks in the network is one of the key techniques,therefore,the research on the classification methods of network attack behavior is particularly important.Classifying network attacks is not a simple task.Due to the diversity and variability of network attacks,many existing classification methods are unable to cope with changing attacks and the emergence of new attacks,and classification accuracy is relatively low.At the same time,overfitting and high bias due to irrelevant or redundant features,as well as the unbalanced class distribution of network attack behavior data,make traditional machine learning methods underperform.In response to the above problems,this thesis builds a model and method for classifying network attack behavior with the help of feature extraction capability of deep neural networks to improve the effectiveness of attack classification and identification of new attacks.The main tasks include:(1)Deep learning is applied to the classification of network attack behavior in response to the complexity and high dimensionality of the raw data type.By comparing different deep learning methods,feature learning of raw network data using autoencoder is proposed.The ability of autoencoder to reconstruct data is used to build depth networks by superimposing autoencoder to obtain low-dimensional deep abstract feature representations of the output data.In view of the slow training speed,poor generalization and overfitting problems in the training process of deep networks,the Re LU activation function is selected and the idea of data normalization and regularization is introduced to optimize the network training process.(2)A cyberattack behavior classification model based on combined feature learning is proposed based on the stacked sparse autoencoder network model.The stacked autoencoder network is first pre-trained to obtain different hierarchical levels of feature representation,then multiple classifiers are trained using these feature sets,and the final classification results are obtained by fusing the decisions of multiple classifiers.The theoretical analysis and experiments show that the proposed method can take full advantage of the features obtained from deep neural network training and can bring the features close to the nature of the data and meet the requirements of the classification task objectives.The constructed classification model shows an improvement in classification performance compared to other methods.(3)Aiming at the situation that the stacked autoencoder network does not fully consider the influence of category features,a network structure based on enhanced category feature learning is proposed.The constraints of category features are added to the training process of the autoencoder network,so that the learned features have relevant information in the same category.On this basis,the influence of unevenly distributed samples is considered,and the training direction of the model is adjusted by improving the loss function,so that the model is more focused on small samples that are difficult to classify.The experimental results show that the center loss within the feature class extracted by the model is reduced,and the feature has a stronger ability to distinguish.While the overall classification effect is improved,the accuracy of discrimination of small sample attack behavior is also improved. |