Framework design of Network intrusion detection based on convolutional neural networks

In view of the limitations of traditional network intrusion detection technologies in dealing with complex attack patterns, this paper proposes a detection framework based on deep learning to improve threat recognition accuracy through automatic feature extraction and nonlinear modeling capabilities. Based on KDD Cup 1999 and UNSW-NB15 data sets, a detection model including convolutional neural network (CNN), long short-term memory network (LSTM) and hybrid architecture was constructed to systematically optimize the data preprocessing process and lightweight model design. Experiments show that the comprehensive performance of the proposed CNN-LSTM fusion model on the two types of data sets is significantly better than that of traditional machine learning and support vector machine models. The detection accuracy of KDD Cup 1999 data set is 90% (F1 value 0.9), and the detection accuracy of UNSW-NB15 data set is 87% (F1 value 0.86). The false positive rate is stable at less than 3%. The experimental results show that the deep learning method significantly enhances the ability to identify covert attacks through multi-level feature abstraction and spatiotemporal correlation modeling, and provides a theoretical basis and technical realization path for building an intelligent network security defense system.