AS-LLM: An LLM-based Framework for Industrial Autonomous System Relationship Inference

Inferring business relationships between Autonomous Systems (ASes) is essential for addressing industrial networking challenges, such as BGP leakage detection. Traditional rule-based methods often suffer from high false-positive rates and limited adaptability, while the potential of modern Large Language Models (LLMs) in this domain remains largely unexplored. In this paper, we introduce the first LLM-driven framework for AS relationship inference, termed AS-LLM. Our methodology tackles the computational limitations of large foundation models, exemplified by DeepSeek-R1, which achieves 80.56% accuracy with 675 billion parameters. We generate high-quality training data from sources like RouteViews, RIPE, and AS-rank to optimize Qwen2.5-7B-Instruct. The resulting refined model, with 7 billion parameters, demonstrates an accuracy of 87.86%, driven by two key innovations: (1) resource-friendly implementation, (2) an 46.31% accuracy improvement through targeted data engineering method.