Linnaeus : A Hierarchical, Multi-Label Framework for Autonomous System Classification Conference: arXiv preprint; ;
Marcos Piotto, Ignacio Schuemer, Santiago T. Torres, Mariano G. Beiró, Esteban Carisimo, Fabián E. Bustamante · 2026
Read Linnaeus : A Hierarchical, Multi-Label Framework for Autonomous System Classification Conference: arXiv preprint; ; by Marcos Piotto, Ignacio Schuemer, Santiago T. Torres, Mariano G. Beiró, Esteban Carisimo, Fabián E. Bustamante (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13649.txt.
Introduction
Marcos Piotto and co-authors present Linnaeus, a system that classifies autonomous systems on the Internet. Such systems play diverse roles, from community backbones to hyperscale content providers and submarine-cable operators. Taxonomies based on network data alone fail to capture how these systems differ in role and operation. Linnaeus adds rich non-network features to topology and BGP data, and uses domain-adapted LLMs with classic machine learning. Its taxonomy has two levels and allows multi-label assignments for hybrid roles. On nearly 2,000 labeled systems it scored well.
Catalog record
| CREATOR | Marcos Piotto, Ignacio Schuemer, Santiago T. Torres, Mariano G. Beiró, Esteban Carisimo, Fabián E. Bustamante |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65752 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.13649
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Linnaeus : A Hierarchical, Multi-Label Framework for Autonomous System Classification Conference: arXiv preprint; ; Marcos Piotto, Ignacio Schuemer, Santiago T. Torres, Mariano G. Beiró, Esteban Carisimo, Fabián E. Bustamante Abstract . Autonomous systems (ASes) play diverse roles in today’s Internet, from community and research backbones to hyperscale content providers and submarine-cable operators. However, existing taxonomies based solely on network-level features fail to capture their semantic and operational heterogeneity. In this paper, we present Linnaeus , a hierarchical AS-classification framework that combines network-centric data (e.g., topology, BGP announcements) with rich non-network features and leverages domain-adapted large language models alongside traditional machine-learning techniques. Linnaeus provides a two-level taxonomy with 18 top-level and 38 second-level classes, supports multi-label assignments to reflect hybrid roles (e.g., research backbone + transit provider), and provides an end-to-end pipeline from data ingestion to label inference. On a manually annotated dataset of nearly 2,000 ASes, Linnaeus achieves an overall precision and recall of 0.83 and 0.76, respectively. We further demonstrate its practical value through case studies, highlighting Linnaeus ’s potential to reveal both structural and semantic dimensions of Internet infrastructure. 1. Introduction The Internet is a vast network of independently operated networks, or Autonomous Systems (ASes). Each AS represents a administrative domain that controls a group of IP prefixes and routing policies. These ASes range widely in scale, ownership, and purpose. They include everything from community networks like NYC Mesh (AS395853) and individual “hobbyist” operators (e.g., AS200556, operated by Ethan Cady) ( 19 ; 39 ) , to hypergiant content providers such as Meta (AS32934) and global transit backbones like Lumen (AS3356). Even small access networks, such as Ilha Turbo in Brazil (AS52803), and isolated deployments serving agricultural facilities (e.g., Ming Yi Tea Farm, AS17415) contribute to the Internet’s decentralized character. As the Internet has evolved beyond its academic roots, researchers have pursued a comprehensive taxonomy of its ASes to answer key questions about Internet structure, resilience, and control. Such classifications help illuminate how limited IPv4 address space is distributed – for instance, highlighting the continued possession of underutilized /8 and /16 blocks by U.S. universities ( 15 ) – and reveal which networks serve as origins for large-scale attacks ( 4 ) or exhibit persistent vulnerabilities to malware ( 31 ; 46 ; 30 ) . Moreover, AS-level classifications support analyses of security adoption (e.g., RPKI deployment across different AS types ( 48 ) ), the role of submarine cable operators in cross-border connectivity, the robustness of government digital infrastructure, and the connectivity choices of critical services like power-grid operators. Ultimately, they reveal structural shifts and increasing centralization, as many formerly self-managed networks migrate to hyperscale cloud providers. Despite decades of effort, a fine-grained and widely accepted taxonomy of ASes remains a challenge. Methods that draw exclusively on network-level data, such as topology snapshots, BGP relationships, or advertised address space, capture structural features yet overlook the semantic context that distinguishes, for example, a government backbone from a cloud platform or a hypergiant CDN from a community network. The Internet’s vast heterogeneity, combined with the lack of consensus on classification criteria, compounds the difficulty. Only a handful of efforts have looked beyond traditional data sources, leveraging business intelligence platforms or crowd-sourced labels, but they still rely on sparse, often incomplete annotations that cannot reflect the full diversity of today’s Internet. Recent breakthroughs in machine learning (ML), particularly domain-adapted large language models (LLMs) capable of combining heterogeneous evidence, open a promising path toward building a richer, more dynamic AS-classification framework. In this paper, we introduce Linnaeus , a hierarchical framework for AS classification. Linnaeus combines state-of-the-art LLMs with traditional ML techniques and leverages both network-centric data and rich, non-network features to generate a two-level taxonomy comprising 18 top-level and 38 second-level classes. In contrast to prior work that assigns each AS a single role, we adopt a multi-label strategy, allowing Linnaeus to capture organizations with multiple operational domains — for example, Internet2 (AS11164) functions simultaneously as a research backbone and a transit provider. This design reconciles the network-oriented and business-oriented paradigms that have historically remained separate. Rather than enforcing mutually exclusive categories, Linnaeus captures the overlapping operational roles that modern ASes routinely play. We demonstrate that Linnaeus is an effective approach. Evaluated on a manually curated dataset of \approx 2,000 ASes, Linnaeus achieves precision and recall of 0.83 and 0.76, respectively. We demonstrate its practical value by classifying the \approx 120,000 allocated ASNs and expanding its value through two case studies: (i) the footprint of government-run networks in the Internet’s network, (ii) the scale of educational and research networks. Our key contributions are: • We introduce an open-source framework for autonomous-system classification. 1 1 1 Available at https://github.com/NU-AquaLab/linnaeus . • By combining complementary machine-learning paradigms, our approach captures multiple operational dimensions of Internet networks more effectively than prior work. • We publish a corpus of nearly 2,000 manually annotated ASes — the largest gold-standard dataset for AS classification to date — and characterize the resulting landscape across all \approx 120k allocated ASNs. This paper does not raise any ethical considerations. 2. Background Despite decades of work, classifying ASes remains a persistent challenge. Since the Internet’s transition to commercial use and today’s ecosystem of hyperscalers, personal ASes, and critical infrastructure networks, researchers have struggled to develop taxonomies that reflect both the functional and organizational diversity of ASes. Early efforts focused on small, manually curated datasets and simple network-centric features. Later approaches introduced richer attributes, such as PeeringDB metadata and customer-cone metrics, but still fell short of capturing ASes’ real-world roles. In this section, we first review key prior efforts (§ 2.1 ), highlighting their contributions and limitations. We then outline how advances in AI and LLMs offer new opportunities to rethink AS classification (§ 2.2 ), enabling richer, more context-aware frameworks like Linnaeus . 2.1. Related Work Several efforts have explored streamlined and automated AS taxonomies over the last two decades. Dimitropoulos et al. ( 16 ) introduced one of the earliest AS taxonomies. By manually inspecting 1,200 ASes, this method proposed six categories, 2 2 2 Large ISPs, small ISPs, customer networks, universities, Internet exchange points, and network information centers. omitting both transit providers and the then-emerging content-delivery networks (CDNs), and a classifying approach solely based on 3 network attributes: organization records, announced IP prefixes, and inferred AS relationships The growing importance of CDNs, with a handful of providers generating the majority of the network traffic, motivates Bottger et al. ( 8 ) to develop an automated identification technique. Leveraging the rich data available in PeeringDB ( 36 ) , including port capacity, geographic footprint, and traffic profile, the authors applied threshold-based heuristics to flag content-provider networks that dominate global traffic. Also building on PeeringDB data, CAIDA ( 13 ) maintained an AS-classification dataset ( 11 ) . The dataset results from combining PeeringDB attributes with the output of a supervised classifier that ingests various features: AS-level degree metrics (customer, provider, peer), customer-cone size, advertised address space, the number of Alexa Top-1M domains served by the AS, and the fraction of its prefixes observed as active by the UCSD Network Telescope. Despite this rich input, the resulting taxonomy included only three classes: Transit/Access, Content, and Enterprise networks. More recently, Ziv