Iot çözümleri için derin öğrenme yöntemlerini kullanan yeni siber güvenlik yaklaşimlari

The emerging Internet of Things (IoT) has transformed the smart environment, industrial systems, and critical infrastructure. However, with this proliferation, numerous challenging cybersecurity issues have emerged due to device diversity, constrained computing capabilities, and a massive flow of data. This thesis explores and proposes new cybersecurity approaches that utilise both classical machine learning and deep learning principles to enhance anomaly detection and robustness in IoT networks. Traditional machine learning models, such as decision trees, support vector machines, and random forests, for anomaly detection in cyber attacks are extensively reviewed in this work. These models are tested on benchmark datasets to establish baseline accuracy and efficiency, particularly in edge scenarios. Based on the above observations, the dissertation presents the design of two deep learning models for IoT security. First, a lightweight FCNN is proposed to detect anomalies in real time on resource-constrained edge devices. Such a model exhibits high accuracy and requires minimal computational effort. Second, we present a hybrid CNN-LSTM network to capture the spatial and temporal features of the IoT traffic. We illustrate the effectiveness of the model by demonstrating its performance on CPIoT-XAD2025, a new cross-domain dataset constructed by combining the N-BaIoT and SWaT datasets in two cyber-physical configurations. All models are then evaluated using accuracy, F1 score, recall, and mean absolute error (MAE). We empirically demonstrate that the proposed architectures achieve better task conditioning and exhibit excellent generalisation across domains. This thesis presents scalable, interpretable and domain-adaptive anomaly detection approaches to address the specific challenges of contemporary IoT systems, providing a basis for future work in intelligent and secure cyber-physical infrastructure.