Hybrid Deep Learning and Boosting Approach for Iot Intrusion Detection Using Autoencoder and Xgboost

The surge in usage of Internet of Things (IoT) infrastructure creates new horizons of security challenges, especially with respect to recognizing malicious activity within networked data. This paper proposes a hybrid multiclass intrusion detection approach for IoT environments, employing both deep learning and machine learning techniques. An autoencoder based on deep learning is employed to compress captured high-dimensional network traffic data to feature representation which summarizes some of the latent patterns of normal behavior in a primary sense. The encoded features are classified using multiple models, specifically XGBoost, Logistic Regression, Naive Bayes and Support Vector Machine (SVM) to classify benign traffic from each of ten attack types. A real dataset (N-BaIoT) consisting of over 1. 8 million labelled records across 9 IoT devices was employed to test the proposed methodology. The experimental results show an overall test accuracy of 9 9. 5 4 \%, with XGBoost demonstrating the best performance of the evaluated classifiers. The results demonstrate the value of using unsupervised feature extraction in conjunction with supervised classification to better develop traceable and accurate intrusion detection models for use in resource-limited environments.