Instructing network devices via large language models

In recent years, there have been significant advancements in Artificial Intelligence (AI), particularly in Natural Language Processing (NLP). These advancements have introduced a groundbreaking approach to effectively manage the escalating complexity of network environments, simplifying network operations by allowing operators to use declarative interfaces instead of imperative ones. This paradigm shift, known as "Intent-Driven Networking," facilitates the conversion of human language into network configurations through the use of NLP techniques, eliminating the necessity for manual coding or execution. Furthermore, the advent of state-of-the-art tools such as GPT, LLama, and PaLM, which possess remarkable capabilities in understanding and generating complex natural language, opens up exciting possibilities for Intent-Driven Networking. This thesis aims to exploit these innovative tools, collectively referred to as Large Language Models (LLMs), by developing a prototype capable of translating high-level policies into actionable network configurations within the realm of Software Defined Networks (SDN). To achieve this objective, we propose a pipeline where the LLM's role is to extract relevant information from human intent. Subsequently, the structured policy extracted is mapped to the Application Programming Interfaces (APIs) of network applications. Through this approach, we want to reduce the gap between high-level policy formulation and network configuration implementation. This model's implementation involves analyzing the few-shot learning capabilities of LLMs to translate a human intent into a machine-readable policies for network devices, utilizing tools such as Langchain to simplify the development of LLMs-based applications and integrating network tools like Ryu-SDN and P4-eBPF. Additionally, the model has been evaluated across various use cases such as firewall, rate-limiting, and load profiling. This evaluation ensures the accurate translation of intent into the appropriate machine-readable format and its effective implementation within the network infrastructure.