AI-Driven Network Slicing for 5G UAV Connectivity Using SDN and Graph Attention Networks in NS-3

Abstract The integration of Unmanned Aerial Vehicles (UAVs) into 5G networks presents a major challenge in ensuring ultra-reliable, low-latency communication (URLLC) under dynamic conditions and diverse user demands. Network slicing, enabled by Software-Defined Networking (SDN) and Network Function Virtualization (NFV), offers a promising solution by creating isolated virtual networks tailored to service requirements. However, static slicing mechanisms struggle to adapt to rapid traffic fluctuations and UAV mobility, resulting in inefficient resource use and degraded Quality of Service (QoS). This paper proposes an AI-driven dynamic network slicing framework to optimize UAV connectivity in a simulated 5G New Radio (NR) environment. To intelligently predict network states and allocate resources, we implement and compare two deep learning models: a Graph Attention Network (GAT) and a Long Short-Term Memory (LSTM) network. GAT has been chosen for its capacity to capture spatio-temporal dependencies in the network graph through attention mechanisms, which illustrate interactions among users, base stations, and UAVs. The framework was validated by conducting extensive simulations in NS-3 using the 5G-LENA module. A multi-slice environment is modeled with diverse traffic types and realistic mobility patterns, including Gauss-Markov for UAVs and Random Waypoint for ground users, across three congestion scenarios. The performance was evaluated based on throughput, latency, packet loss, and resource utilization. Results show that the GAT-based framework consistently outperforms the LSTM baseline, achieving superior congestion prediction, reduced latency, and improved resource allocation. These findings highlight the role of AI-enhanced slicing in meeting UAV application demands and provide a foundation for intelligent resource management in next-generation wireless networks.