Intrusion detection system based on machine and deep learning approaches

In this thesis, three different intrusion detection systems (IDS) named PBCSC-IoT, PMBL-IoT, and AOCBiT were developed for Internet of Things (IoT) networks, and their performances were thoroughly analyzed and compared. The effects of the feature reduction, data balancing, and classification methods used in the proposed IDS models on IDS performance were evaluated, and the detection success of these models was compared in detail. The PBCSC-IoT model was created using a hybrid approach that combines the Bat Optimization Algorithm (BAT), Principal Component Analysis (PCA), Synthetic Minority Oversampling Technique (SMOTE), and Convolutional Neural Network (CNN) methods, and was tested on the IoTID20 and BoT-IoT datasets. The PMBL-IoT model was developed using PCA, Mayfly Optimization Algorithm (MAO), Borderline SMOTE (BSMOTE), and Long Short-Term Memory (LSTM) methods and was tested on a combined dataset obtained by merging the IoTID20, CIC-ToN-IoT, and USB-IDS-1 datasets. The AOCBiT-IoT model was developed using an Autoencoder, Analysis of Variance F-Test (ANOVA F-Test), Adaptive Synthetic Sampling (ADASYN), One-Sided Selection (OSS) method along with CNN, Bidirectional LSTM (BiLSTM), and Transformer methods, and was tested on the CIC-DDoS2019 dataset. The results demonstrated that hybrid deep learning techniques provide high performance in detecting attacks in IoT networks and that data balancing and feature reduction methods can further enhance this classification performance.