Deep learning for network security: an Attention-CNN-LSTM model for accurate intrusion detection
Abdullah Mujawib Alashjaee
Abstract
Intrusion Detection Systems (IDS) are vital for protecting networks with evolving cyber threats, that comprises malware, denial-of-service attacks, and botnets. Hence, this study proposes a novel hybrid deep learning model, which associates Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), along with a self-attention method to highlight the utmost informative input features. Here, CNN works with extraction of spatial features, LSTM works with modelling of temporal sequence. The proposed model is termed Attention-CNN-LSTM. The proposed model achieves 94.8-97.5% accuracy and significantly improves Matthews Correlation Coefficient (MCC) and F1-score by evaluating on NSL-KDD and Bot-IoT datasets. An ablation study confirms each component contribution, particularly the attention layer, to overall performance gains. The architecture supports real-time inference with sub-35ms latency. The model also shows strong potential for real-time deployment, processing over 1200 records per second; hence, this work is applicable for high-traffic environments.