Unleashing AI-Empowered Slices on Mobile Networks for Natively Cognitive Service Delivery

Integrating Artificial Intelligence (AI) into network slice orchestration remains a challenge due to data availability, model generalization, and the need for real-time inference. The Network Data Analytics Function (NWDAF) systematically incorporates AI into the modern mobile network core to enhance network operation management. Although research fronts have explored AI applications in mobile networks, primarily to improve performance or connectivity management, orchestrating and delivering AI-driven slices for end users still requires resolving issues of scalability, interoperability, and latency. The literature categorizes these challenges into applying AI for autonomous network management, which involves a mature and evolving orchestration to support stringent AI-based slices for end-users. This work addresses the latter, proposing enhancements to two core network components, the User Plane Function (UPF) and NWDAF, to enable seamless deployment of AI slices for both training and inference targeted at mobile users. We introduce, for the first time, distinct training and inference slices orchestrated through an enhanced UPF (e-UPF) and NWDAF within the 5G core, where the UPF performs slice-aware forwarding of AI traffic to Multi-access Edge Computing (MEC) servers. All AI training and inference workloads are executed at the network edge (MEC). With this approach, the core network can deliver cognitive services directly to User Equipments (UEs). A disaster management case study demonstrates this, where Convolutional Neural Networks (CNNs) deployed on MEC servers classify disaster scenarios via AI slices. Results show that AI slices can be efficiently deployed and differentiated via data plane filters, enabling private cognitive services and paving the way for AI-based applications in mobile networks.