احمد جبار قاسم الخماسي

The rapid growth of the Internet of Things presents significant security challenges, particularly in critical infrastructures, industries, and smart cities. This thesis investigates the application of deep learning models to improve intrusion detection systems in Internet of Things networks, focusing on enhancing detection accuracy and computational efficiency through hyperparameter tuning. Four deep learning classifiers were eva‎luated: Radial Basis Function Network, Long Short-Term Memory Network, Convolutional Neural Network, and Multilayer Perceptron. These models were trained on the CIC-IoT2023 dataset for binary classification, multi-class classification of eight attack types, and classification of 33 specific attacks. A 10-fold cross-validation approach was used, measuring performance by F1 score, precision, recall, training time, and prediction time. Our Resluts indicate that the Convolutional Neural Network outperformed the other models, achieving the highest F1 score of 97.4%, with strong precision 97.2% and recall 97.6%, though it required longer training and prediction times. The Long Short-Term Memory Network also performed well, especially in recall 98.4%, but its extended training time 518.34s and prediction time 8.45s could hinder real-time applications. The Radial Basis Function Network had the fastest prediction time 3.12s but suffered from low precision 81.1%, resulting in a lower F1 score 88.8%, making it less reliable for high-risk environments. The Multilayer Perceptron offered a balanced approach, achieving an F1 score of 93.3% with relatively short training and prediction times, making it suitable for computationally efficient scenarios. This study underscores the importance of hyperparameter tuning in deep learning-based intrusion detection systems for Internet of Things environments. The Convolutional Neural Network is highlighted as the most reliable for real-time security applications, while the Long Short-Term Memory Network provides an optimal balance between detection accuracy and speed. The Radial Basis Function Network requires further optimization to be viable for high-stakes intrusion detection, while the Multilayer Perceptron serves as a computationally efficient alternative.

امنيت سايبري , سيستم‌ تشخيص نفوذ , اينترنت اشيا , مجموعه داده سي‌آي‌سي (2023) , يادگيري عميق , تنظيم فراپارامتر , حافظه كوتاه - بلند مدت , شبكه عصبي كانولوشنال , شبكه تابع پايه شعاعي , مدل پرسپترون چندلايه

Cyber Security , Intrusion Detection System , Internet of Things , CIC-IoT2023 Dataset , Deep Learning , Hyperparameter tuning , Short-Long-Term Memory , Convolutional Neural Network , Radial Basis Function Network , Multilayer Perceptron Model

Ahmed Al-Khammasi

Dr. Marzieh MalekiMajd