TITLE:
Deep Learning-Based Two-Step Approach for Intrusion Detection in Networks
AUTHORS:
Kamagaté Beman Hamidja, Kanga Koffi, Kouassi Adless, Olivier Asseu, Souleymane Oumtanaga
KEYWORDS:
Cybersecurity, CICIDDS2017, Intrusion Detection, BiLSTM, Deep Auto-Encoder
JOURNAL NAME:
International Journal of Internet and Distributed Systems,
Vol.6 No.2,
November
22,
2024
ABSTRACT: Intrusion Detection Systems (IDS) are essential for computer security, with various techniques developed over time. However, many of these methods suffer from high false positive rates. To address this, we propose an approach utilizing Recurrent Neural Networks (RNN). Our method starts by reducing the dataset’s dimensionality using a Deep Auto-Encoder (DAE), followed by intrusion detection through a Bidirectional Long Short-Term Memory (BiLSTM) network. The proposed DAE-BiLSTM model outperforms Random Forest, AdaBoost, and standard BiLSTM models, achieving an accuracy of 0.97, a recall of 0.95, and an AUC of 0.93. Although BiLSTM is slightly less effective than DAE-BiLSTM, both RNN-based models outperform AdaBoost and Random Forest. ROC curves show that DAE-BiLSTM is the most effective, demonstrating strong detection capabilities with a low false positive rate. While AdaBoost performs well, it is less effective than RNN models but still surpasses Random Forest.