Multi-Class Network Intrusion Detection Using Deep Neural Networks Tuned on Imbalanced Data
Shruti Tyagi, Shriya Pingulkar, Amaan Shaikh · 2023
As computer networks become ubiquitous, they face escalating risks from cyberattacks that can severely disrupt critical infrastructure. Existing network intrusion detection datasets have limitations like inadequate diversity of samples and poor classification accuracy. This research utilizes the University of Nevada - Reno Intrusion Detection Dataset (UNR-IDD) encompassing normal traffic and 5 different intrusion types to develop a deep neural network for multi-class intrusion detection. After exploratory data analysis, a baseline Multi-Layer Perceptron model was developed. A structured tuning methodology involving ANOVA feature selection and grid search hyperparameter optimization is followed to improve validation performance. The research quantitatively demonstrates the impact of methodical tuning guided by validation accuracy, establishing a template for developing high-performance deep learning intrusion detection systems. With further enhancements, the techniques explored could equip security experts with invaluable intelligence to protect networks against sophisticated cyber threats