A LIGHTWEIGHT HYBRID DEEP LEARNING FRAMEWORK FOR MULTI-CLASS NETWORK INTRUSION DETECTION

Abstract

The growing variety of cyberattacks has increased the need for intrusion detection systems that can detect multiple categories of attack with high reliability. We propose a lightweight hybrid framework LHCBA-Net, which integrates one dimensional convolutional neural networks, bidirectional long short term memory and temporal attention for multiclass network intrusion detection in this paper. The model was trained after the CICIDS2017 dataset went through a series of data cleaning, feature engineering, feature selection, class balancing and feature transformation procedures. CNN layers capture local traffic patterns, BiLSTM models bidirectional contextual dependencies, and temporal attention highlights informative intrusion related representations. Experimental results on the test set compared with traditional machine learning and deep learning baselines show that LHCBA-Net achieved 99.08% test accuracy with precision, recall, F1-score and AUC of 0.991, 0.990, 0.990 and 0.998 respectively. Feature-importance and SHAP analyzes provide further support for the interpretation of the learned decision patterns. The results show that the proposed framework is efficient for accurate detection of multi-class intrusions.