Network Intrusion Detection with SMOTE-ENN and Deep Learning Techniques

Background/Objectives: Intrusion detection is paramount for protecting modern information systems from cyber threats. However, the task faces significant challenges stemming from imbalance class distribution, high dimensionality, complex network traffic and high resource constraints. This paper aims to address these hurdles by proposing a novel approach to Network Intrusion Detection Systems (NIDS). Methods: Leveraging advanced machine learning techniques, our model employs a two-step resampling technique, SMOTE-ENN, to tackle imbalanced datasets and mitigate noise. Additionally, the model uses parallel autoencoders for normal and attack classes to perform feature extraction and dimensionality reduction. It then utilizes a lightweight Multi-Layer Perceptron (MLP) classifier to differentiate between normal and abnormal network activities. Findings: Experimental evaluations on benchmark datasets, including UNSW-NB15 and NSL-KDD, demonstrate the effectiveness of our model. It outperforms traditional methods, achieving superior performance in real-world cybersecurity scenarios characterized by imbalanced network traffic. Novelty: Our proposed approach presents a novel integration of advanced deep learning-based techniques aimed at improving the accuracy and efficiency of intrusion detection systems. It offers a lightweight and robust solution for detecting specific attacks within network data.