Dynamic CNN-LSTM-Based Intrusion Detection System in Enterprise Financial Systems

In traditional network intrusion detection research, the feature selection and extraction process is time-consuming and prone to bias. The detection effect on novel attacks such as zero-day attacks and distributed denial of service attacks is poor, and there needs to be more dynamic learning and adaptation mechanisms, resulting in low detection accuracy and processing capabilities. Therefore, this paper applies deep neural network (DNN) technology and uses convolutional neural networks (CNN) to automatically extract the spatial features of network traffic data to enhance the ability to recognize attack patterns. Meanwhile, long short-term memory (LSTM) networks are used to capture time series features and identify time series changes in attack behaviors. Second, the model is continuously trained and updated through incremental methods to enhance its adaptability to novel attacks. The system uses a dynamic learning mechanism based on the feedback mechanism and data augmentation to monitor and optimize the detection effect in real-time. Compared with the support vector machine (SVM) model and the CNN model, the CNN-LSTM model in this paper performs better than the other two, with the highest detection accuracy of 98.6%, the highest processing speed of 3440.15[Formula: see text]kB/s, the highest false positive rate of 0.85% and the highest false negative rate of 0.6%. The experimental results suggest that the intrusion detection system based on DNN built in this paper has advantages in identifying network attacks and provides effective technical support for the security protection of enterprise financial systems.