The Politics of Routing: Investigating the Relationship between AS Connectivity and Internet Freedom

Rachee Singh, Hyungjoon Koo, Najmehalsadat Miramirkhani, Fahimeh Mirhaj, Phillipa Gill, Leman Akoglu

TL;DR

Researchers from Stony Brook University have demonstrated that a country’s level of Internet freedom is physically etched into its network topology. By combining network measurement data with machine learning, they developed a model that predicts a nation's "Internet Freedom" category with 91% accuracy based on features like IP density and routing path lengths.

What Problem Does This Paper Address?

Measuring "Internet Freedom" is traditionally a manual, qualitative process. Organizations like Freedom House produce annual reports based on expert analysis of laws, censorship, and crackdowns. However, this approach is labor-intensive and often suffers from data gaps in regions where measurement points are scarce.

The authors of this paper ask a provocative question: Does a country's political stance on information control leave a measurable footprint on its physical Internet infrastructure? For example, if a government mandates that all traffic must pass through a single state-controlled gateway (like in Iran), that policy fundamentally alters the graph structure of the country's networks. This paper aims to quantify that relationship and use network topology as a proxy for freedom.

Methods

The researchers combined network measurement techniques with machine learning to bridge the gap between "bits and ballots."

  1. Topology Mapping: They started with the CAIDA Autonomous System (AS) relationship dataset. To fix known "blind spots" in this data, they used the RIPE Atlas platform to perform thousands of targeted traceroutes (Inside-out, Outside-in, and Mesh) to uncover nearly 6,000 previously hidden domestic and international network edges.
  2. Feature Engineering: They extracted 20 distinct features from these network graphs, categorized into:
  3. Structural: Graph metrics like transitivity (the number of "triangles" or redundant paths) and centrality.
  4. International Connectivity: The number of countries a nation connects to and the length of those paths.
  5. IP Demographics: Specifically, IP Density (IP addresses per capita).
  6. BGP Routing: Features related to ISP properties and "customer cone" sizes.
  7. Machine Learning: They tested several models, finding that a Decision Tree with Linear Regression at the leaves (DTLR) performed best. This model groups similar countries together and then applies a specific regression formula to each group to predict their Freedom of the Press Index (FPI) score.

Datasets

  • CAIDA AS Relationship Dataset: The baseline for the global Internet map.
  • RIPE Atlas: A global network of probes used to perform active measurements and find missing network links.
  • Freedom House Freedom of the Press Index (FPI): Used as the "ground truth" for 170 countries. They chose this over the "Freedom on the Net" score because it covers significantly more countries (199 vs. 66).
  • Regional Internet Registries (RIRs): Used to map specific ASes to their respective countries.

Key Insights and Ideas

  • IP Density is the "Golden Feature": The single most important predictor of freedom was IP density. A high ratio of IP addresses to the population strongly correlates with high freedom of expression. This likely reflects both economic development and a lack of centralized IP management used for control.
  • Path Length as a Control Mechanism: There is a negative correlation between freedom and the maximum length of BGP-compliant paths to other countries. In simpler terms: the more "hops" it takes to get out of a country, the less free that country tends to be. Longer paths often imply more centralized chokepoints where filtering can occur.
  • The Infrastructure of Censorship: Restrictive regimes often have topologies with low transitivity (fewer redundant connections between ISPs) and high centrality, making it easier to implement country-wide shutdowns or deep-packet inspection.
  • Identifying "Unusual" Countries: The model highlighted outliers like Singapore and Botswana. Singapore has a highly connected, "free-looking" topology but a low freedom score due to strict media laws. Conversely, Botswana has high freedom but a "restricted-looking" topology, likely due to infrastructure development lagging behind its political progress.

Key Results

  • The DTLR model achieved an average prediction error of only 7.06% when estimating a country's FPI score.
  • The model predicted the specific freedom category (Free, Partly-Free, or Not Free) with 81% accuracy.
  • When the task was simplified to distinguishing between "Free/Partly-Free" and "Not Free," the accuracy climbed to 91%.

Limitations and Future Work

The authors acknowledge that the AS-level graph is still an approximation; even with RIPE Atlas, some peering relationships remain hidden. Furthermore, the study identifies correlation, not necessarily causality. It remains a "chicken and egg" problem: does a centralized topology enable censorship, or does a censorial government mandate a centralized topology?

Future work aims to incorporate temporal analysis—tracking how a country's network structure evolves as its political climate changes—and collaborating with social scientists to better understand the policy drivers behind these physical network configurations.

Summary updated

Abstract

The Internet's importance for promoting free and open communication has led to widespread crackdowns on its use in countries around the world. In this study, we investigate the relationship between national policies around freedom of speech and Internet topology in countries around the world. We combine techniques from network measurement and machine learning to identify features of Internet structure at the national level that are the best indicators of a country's level of freedom. We find that IP density and path lengths to other countries are the best indicators of a country's freedom. We also find that our methods predict freedom categories for countries with 91% accuracy.
BibTeX
@inproceedings{singh2016politics,
  author = {Rachee Singh and Hyungjoon Koo and Najmehalsadat Miramirkhani and Fahimeh Mirhaj and Phillipa Gill and Leman Akoglu},
  title = {{The Politics of Routing: Investigating the Relationship between AS Connectivity and Internet Freedom}},
  booktitle = {FOCI},
  year = {2016},
  url = {https://openreview.net/forum?id=PS6Vxyi3gvT},
}

Loading related content...

Figure