Affinity-Model: Improving AS Routing Models via AS Affinity Behavior Inference

Understanding the routing behavior exhibited by ASes is crucial for us to comprehend the Internet workings, optimize diverse network applications, and facilitate the evolution of the Internet. However, prior research often models AS routing behaviors with oversimplified AS relationships, lacking explanations for complex behaviors, thus hampering their own accuracy. In this paper, we introduce Affinity-Model (AM), a novel approach designed to discover AS affinity behaviors that traditional assumptions cannot explain, thereby enhancing existing routing models. AM first uses the state-of-the-art AS-level path inference algorithm to infer the RIB-in path of each source AS. Discrepancies between the observed routing selection and the optimal path determined by traditional sorting methods then identify the AS affinity behaviors. Second, AM leverages the high-occurrence affinities discovered by full VPs, extracts their key features, and makes extensive inferences about remaining affinities across the entire Internet. Third, AM extends its utility to infer non valley-free policies in the Internet. Findings indicate a strong association between affinities and AS prefix/destination-specific policies, with a higher incidence in European IXP connections. Our experiments show that the inferred AS affinity behaviors align highly with IRR-recorded policies, modeling 85.9% of AS routing behaviors in the actual Internet.