A research project that reimagines the mobile core as both a consumer and a provider of Artificial Intelligence. Net-AI-APPs evolves the 3GPP Network Data Analytics Function (NWDAF) and the User Plane Function (UPF) into an open, microservice-based platform where network functions, operators, and third-party B5G applications can request, train, and consume AI services — turning raw network metrics into actionable, cognitive, and sustainable services.
From B2B-driven 5G to a natively cognitive Beyond-5G core.
5G was designed with a strong B2B focus, delivering gains in latency, data rate, and reliability across industry verticals. To meet the specific needs of each vertical, ETSI introduced the concept of the NetApp — a virtual, 5G-enabled application that can operate autonomously and cooperate with other NetApps to compose more complex vertical services. In the 5G Core, the NWDAF is the function responsible for processing network data and producing insights, but it is traditionally managed exclusively by the operator, limiting external developers and integrators.
Net-AI-APPs proposes an enhanced, microservice-based architecture for the NWDAF that lets external entities — the NetApps — consume existing analytics or create new ones, adding new analytic and model-training functions on demand. This introduces the notion of Network Metrics as a Service (NMaaS), exposing AI metrics of the mobile network in a distributed and accessible way, and creating the conditions for a new class of AI-native applications: the Net-AI-APPs.
Scientific and technological results directly tied to the evolution of AI-driven B5G networks.
An enhanced NWDAF and UPF architecture that delivers AI-specific slices for training and inference, routing traffic to MEC resources via TEID. Implemented on free5GC and validated in a disaster-classification use case.
DOI ↗High-resolution UPF observability built with eBPF/bpftrace, measuring per-packet GTP-U decapsulation latency and quantifying noisy-neighbor interference between slices in the B5G data plane.
DOI ↗A closed-loop, agentic control architecture where local LLMs consult telemetry and execute verifiable routing actions on the UPF, translating high-level intents into energy-aware network operations.
DOI ↗Predictive, telemetry-driven autoscaling for container-based services that cut energy consumption by 34.68% versus the Kubernetes HPA baseline while preserving latency targets.
DOI ↗Federated semantic segmentation of synthetic UAV imagery under non-IID conditions, combining HRNet, FedAvg, and a composite loss to study convergence, communication, and per-class performance.
DOI ↗A traceroute-based approach using temporal feature engineering and ensemble learning to detect rare Internet route changes without access to the control plane.
DOI ↗Researchers and scholarship students who built Net-AI-APPs.
A quick, plain-language walkthrough of how the mobile network can understand itself and deliver AI services.