Cyber intrusion detection using ensemble of deep learning with prediction scoring based optimized feature sets for IOT networks

Abstract

Detecting intrusions in Internet of Things (IoT) networks is critical for maintaining cybersecurity. Traditional Intrusion Detection Systems (IDS) often face challenges in identifying unknown attacks and tend to have high false positive rates. To address these issues, we propose the Ensemble of Deep Learning Models with Prediction Scoring-based Optimized Feature Sets (EDLM-PSOFS). Our approach begins with data preprocessing utilizing MissForest imputation and label one-hot encoding, effectively managing incomplete and categorical data.

For feature selection, we employ the Median-based Shapiro-Wilk test alongside Correlation-Adaptive LASSO Regression (CALR) to ensure robust feature extraction. To capture temporal patterns effectively, our ensemble integrates Global Attention Long Short-Term Memory networks (GA-LSTMs), utilizing layered structures, residual connections, and attention mechanisms. Additionally, to enhance interpretability and support decision-making, we incorporate the Exploit Prediction Scoring System (EPSS), which evaluates prediction scores and provides detailed insights, thereby improving overall model performance. This comprehensive methodology aims to strengthen the detection capabilities of IDS in IoT environments, reducing false positives while effectively identifying unknown threats.