MSCNN-BiLSTM: A Network Intrusion Detection Model Optimized by Genetic Algorithm with Deep Spatio-Temporal Feature Learning

Abstract

Against the backdrop of accelerated digital transformation and evolving cyberattacks, traditional intrusion detection methods face severe challenges due to their limited feature extraction capabilities, fragmented spatiotemporal feature processing, and insufficient model adaptability. These challenges include a high proportion of encrypted traffic and the emergence of novel attacks. To enhance detection accuracy and adaptability, this paper proposes a Multi-Scale Convolutional Neural Network-Bidirectional Long Short-Term Memory Network (MSCNN-BiLSTM) model optimized by genetic algorithms. This approach utilizes the multi-branch structure of MSCNN (with 1x1, 3x3, and 5x5 convolutional kernels) coupled with a spatial attention mechanism to strengthen spatial feature extraction. It combines BiLSTM and a self-attention mechanism to capture temporal dependencies, and employs genetic algorithms to automatically optimize hyperparameters, achieving efficient synchronous learning of spatiotemporal features in network traffic. Experimental results on the UNSW-NB15 dataset demonstrate that the model achieves an accuracy of 96.8%, a recall rate of 96.4%, a precision rate of 97.2%, and an F1-score of 96.7%. In robustness tests, the performance loss under noise and adversarial attacks is controllable (average <5%). The average performance loss in cross-scenario transfer experiments is only 4.2%, significantly outperforming comparative models. The model proposed in this paper effectively addresses the limitations of traditional methods in feature extraction and adaptability. Its main contribution lies in the innovative integration of a deep spatiotemporal feature learning framework, providing key technical support for building a highly accurate and strongly adaptive network intrusion detection system. This has significant theoretical value and broad application prospects.