EN-US">The advent of fifth-generation (5G) networks and the upcoming sixth generation (6G) introduces unprecedented complexity in managing heterogeneous and dynamic network infrastructures. To address this, the Intent-Based Networking (IBN) paradigm allows users to express high-level objectives, or intents, which are automatically translated and applied by the network. However, current IBN solutions remain limited by their reliance on rigid specifications (e.g., JSON, YAML) and constrained autonomous assurance mechanisms, making intents difficult for non-expert users to formulate and limiting the continuous alignment between intended goals and actual network behavior.
EN-US">This thesis investigates how Generative Artificial Intelligence (GenAI), particularly Large Language Models (LLMs), can serve as a key enabler for next-generation IBN architectures. LLMs have demonstrated remarkable capabilities in natural language understanding, reasoning, and adaptability, making them suitable to bridge the gap between human-centered specification and machine-level execution. Three major challenges are addressed: (i) intent translation, (ii) intent assurance, and (iii) GenAI operations in IBN.
EN-US">For intent translation, we propose LLM-based architectures that allow intents to be expressed in natural language and automatically translated into deployable network configurations. We focus on network service configuration across multiple technological domains, including the RAN, CN, and Edge/Cloud. LLMs effectively convert user-friendly intents into service descriptors, lowering the barrier for non-expert users. This requires decomposing complex intents into coordinated, domain-specific tasks while managing the full intent lifecycle, including negotiation, configuration, monitoring, and assurance. We further extend translation to interactions with OSSs, where a conversational agent plans and executes multiple API calls to satisfy high-level intents in heterogeneous environments.
EN-US">For intent assurance, we demonstrate how GenAI can ensure that deployed services continuously meet defined objectives. We design a modular architecture based on the NWDAF network function, incorporating ML models for anomaly detection in UE traffic, validated on real experimental environments. We then propose a closed-loop assurance pipeline integrating AI, eXplainable AI (XAI), and LLMs, capable of autonomously detecting and resolving anomalies while explaining root causes and corrective actions in natural language, enhancing operator transparency and trust for Zero-touch network and Service Management (ZSM).
EN-US">Finally, for operational deployment, we introduce 5G INSTRUCT Forge, a data-processing pipeline that systematically extracts knowledge from 3GPP technical specifications to create Telecom-specialized LLMs, outperforming general-purpose models on domain-specific tasks. In addition, a Deep Reinforcement Learning (DRL) scheduler was proposed to dynamically allocates tasks to LLMs under strict SLO constraints, ensuring scalability and efficiency with limited AI resources.
EN-US">The proposed architectures were validated through prototypes and extensive 5G/6G experimental deployments, including EURECOM’s 5G facility. Results show significant improvements in ease of intent specification, translation accuracy, assurance reliability, interpretability, and resource efficiency compared to existing approaches.
EN-US">In summary, this thesis establishes GenAI as a key technology for IBN in 6G, enabling natural-language intent specification, reliable autonomous assurance, and scalable AI operations. It paves the way for user-centric, autonomous, and resilient 6G networks, while opening new research directions in federated intent management, responsible AI, and multi-operator collaboration.
