Artificial intelligence-assisted detection and forensic analysis of cyber threats to unmanned aerial vehicles

In today's rapidly accelerating digitalization era, Unmanned Aerial Vehicles (UAVs) have become widely used across civilian, industrial, and military domains. With features such as autonomous mission execution, real-time data collection, and flexible operational capabilities, these systems stand out while simultaneously exposing a new and vulnerable surface to cybersecurity threats. This expanded attack surface has made UAV systems high-value targets for cyber attackers. In this context, the aim of this thesis is to simulate various cyberattacks targeting UAVs, analyze their effects on flight parameters, and evaluate the detectability of such attacks using artificial intelligence-based methods. Using the PX4 flight stack, cyberattack scenarios including GPS Spoofing, Control Hijack, and Velocity DoS were developed in jMAVSim and Gazebo simulation environments. After each flight, the resulting log files in .ulg format were converted into CSV format for analysis. Data preprocessing steps such as cleaning, feature selection, standardization, and labeling were performed, and the attacks were modeled in a multi-class classification structure. During the modeling phase, algorithms such as XGBoost, LightGBM, Random Forest, and Logistic Regression were employed. According to the obtained results, LightGBM achieved the highest performance with 96.3% accuracy and 95.8% F1-score. Further evaluations using ROC AUC and Confusion Matrix analyses confirmed the consistent classification success of the LightGBM model.