Enhanced Network Metric Prediction for Machine Learning-Based Cyber Security of a Software-Defined UAV Relay Network

Unmanned aerial vehicles (UAVs), e.g., drones, have become crucial assets in the military’s fleet of vehicles. UAVs can provide limited bandwidth for tactical communications and can act as relays over battlefields. Modern drones provide much higher bandwidth than their legacy predecessors with dynamic antenna capabilities that would be useful in communicating around obstacles, such as urban corridors formed by rows of tall buildings that limit terrestrial lines of sight and attenuate high frequencies. While it is still more likely that one UAV is used for this purpose, a well-managed cluster of UAVs could increase the functionality of the entire terrestrial-drone network. Software-defined networking (SDN) is recognized as an effective way to manage distributed wireless networks. In our previous research, we used a software-defined UAV (SD-UAV) network to offer organized and secure communication resources and relaying capabilities to soldiers, military vehicles, and assets on the ground in a signal-challenged urban environment. Additionally, we developed a framework for detecting common multiple types of cyberattacks (black hole, gray hole, and jamming) using the Light Gradient Boosting machine learning algorithm. However, zero-day attacks are newly developed cyberattacks that can cause significant harm to the functionality of a SD-UAV network. Due to the novelty of these attacks, there is a lack of data samples for ML model training and testing. In order to tackle this problem, this paper provides a predictive queuing analysis of a SD-UAV network. To our knowledge, no previous work has proposed using queuing analysis to predict network performance statistics of SD-UAV networks for the purpose of enhancing security against zero-day cyberattacks. This analysis allows the network operator to calculate the expected theoretical values of the average interarrival time, transmission delay, and packet count of each SD-UAV in a swarm under normal, non-attack operating conditions. This allows the network operator to establish a normal baseline for ML training in order to compare and detect zero-day attacks by detecting any malicious or abnormal SD-UAV network behavior. In this study, we conduct an analysis, simulation, and discussion of the results of our predictive queuing analysis. Our simulation results confirm the accuracy of our predictive queuing analysis we obtained for the average interarrival times, transmission delay, and packet count for the SD-UAVs in the swarm. Our predictive queuing analysis enables the network operator to predict network performance statistics of the SD-UAVs for the purpose of enhancing security against zero-day cyberattacks.