With the rapid growth of network technologies and internet-based services, cybersecurity threats and network intrusions have become major challenges for modern information systems. Traditional intrusion detection systems often face limitations in accurately detecting complex and evolving cyberattacks, particularly when handling large-scale network traffic data. Therefore, the use of deep learning techniques has attracted significant attention due to their capability to automatically learn representative patterns from network traffic data. This thesis presents a deep learning-based intrusion detection framework for detecting and classifying malicious network traffic using the NSL-KDD dataset. Several preprocessing operations were applied before model training, including categorical feature encoding, feature normalization, dataset reshaping, and train-validation-test splitting in order to prepare the dataset for deep learning analysis. The proposed framework implemented and evaluated four different deep learning architectures, including Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN-1D), Bidirectional Long Short-Term Memory (BiLSTM), and Transformer models. The performance of the implemented models was evaluated using Accuracy, Precision, Recall, and F1-score metrics in order to provide a comprehensive assessment of intrusion detection performance. The experimental results demonstrated noticeable differences among the implemented architectures. The MLP model achieved the strongest overall classification performance with an accuracy of 79.64% and an F1-score of 78.29%, outperforming the CNN-1D, BiLSTM, and Transformer models on the NSL-KDD dataset. The obtained findings indicated that simpler dense learning architectures were more suitable for the tabular structure of the dataset than more complex recurrent and attention-based models. The study also revealed that dataset imbalance significantly affected the classification performance of minority intrusion categories such as R2L and U2R attacks. Despite the application of regularization and optimization techniques such as dropout and early stopping, minority attack categories continued to produce lower Recall and F1-score values across most implemented models. Overall, the proposed framework demonstrated the capability of deep learning techniques to support intelligent intrusion detection systems in cybersecurity environments. The findings of this study provide useful insights into the behavior of different deep learning architectures for intrusion detection applications and highlight the importance of selecting suitable architectures according to dataset characteristics and cybersecurity requirements.
