Artificial intelligence (AI) methods and tools for improving network monitoring practices

The Internet monitoring infrastructure provided publicly by RIPE NCC, RouteViews, and other organizations, is invaluable for network operators, who use it to monitor their Internet routes, detect routing events that affect their traffic, troubleshoot their networks, discover the topology and design their policies, etc.

However, it is well known that the current infrastructure is not uniformly deployed, and this may limit some measurement efforts (e.g., regions with poor coverage) or lead to biased findings (e.g,. over/under-estimate impact of a route event depending on location).

The goal of the AI4NetMon project is to study limitations and biases, and develop tools and strategies for improving the Internet monitoring infrastructure and practices, by addressing two key questions:

Do you use the Internet monitoring and measurement infrastructure? Participate in our online survey, and help us to identify and fix biases!


To answer the research questions AI4NetMon will be based on a data-driven approach, leveraging on rich heterogeneous network data and state-of-the-art ML/AI methods.

The project will develop a set of open-source tools, aimed to be used by Internet operators, researchers, and practitioners (without requiring expertise in ML/AI). The AI4NetMon tools will receive as input the available network data and provide as output answers to the key questions (e.g., potential locations for new monitor deployments, monitoring selection strategies, unbiased estimates for a set of measurements).

Specifically the project objectives are to:

Identify the key use cases for monitoring infrastructure

Which are the best locations to deploy monitors? What is the best set of probes to conduct measurements from? ...There is not a single answer to these questions! The best measurement strategy depends on what one wants to measure; e.g., a best set of monitors for detecting an outage, may be different than the best set of monitors for estimating an anycast catchment. Since a global answer does not exist, we need to provide answers per case. AI4NetMon will identify the use cases of higher importance to network operators, and for which the monitoring infrastructure is more frequently used.

Quantify the bias and efficiency of the existing infrastructure and potential for improvement

What is the accuracy, coverage, and bias of the existing monitoring infrastructure (for a certain use case)? What would be the best achievable performance? Is there a margin for improvement? How much? Building upon preliminary findings (paper), AI4NetMon aims to answer these questions through theory, measurements, and experiments.

Develop open-source tools for efficient Internet monitoring

AI4NetMon will develop AI methods and tools to (i) identify the best locations (e.g., ASes, IXPs, geographical regions) to deploy new infrastructure, (ii) select set of vantage points from the existing infrastructure for lightweight and efficient measurements (e.g., subset of Atlas probes or route collectors), and (iii) unbias estimations from existing measurements. The AI4NetMon tools will collect data from various heterogeneous data sources and aggregate them in a common dataframe, which will feed state-of-the-art AI that will provide recommendations, predictions, inferences, etc., for the desired monitoring objectives.

Project results

Bias: Quantification, Metrics, Results

We quantify the bias in the Internet measurement/monitoring infrastructure and platforms. We first define the bias and the metrics to quantify it. Then we compare the bias in the different Internet measurement infrastructures (e.g., RIPE Atlas, RIPE RIS, RouteViews) and present results that visualize this bias. The answers to the following questions

  • How biased is the infrastructure? (e.g., from a scale from 0 to 1, where 0 is no bias)
  • Is RIPE RIS more biased than RIPE Atlas? If yes, how much? What about RouteViews?
  • Is the infrastructure more biased in terms of location or network sizes?
can be found in the detailed page about the bias in the Interent monitoring infrastructure and in our RIPE Labs article.

image 1
Radar plot depicting the bias of the RIPE Atlas and RIPE RIS infrastructures along different dimensions (network size, topology, location, etc.)

Recommendations for deployments of extra RIPE RIS and RIPE Atlas vantage points

We aim to find new potential locations for RIPE RIS monitors (i.e., ASes to peer with RIPE RIS). When adding an AS to the RIPE RIS monitors, then (i) the bias of the RIPE RIS changes (can increase or decrease), and (ii) there is some "improvement score" in terms of proximity to origin ASes, i.e., the total distance in AS hops to origin ASes decreases).

The following plot shows what is the trade-off by adding an AS to RIPE RIS: bias difference by adding this AS to RIPE RIS (x-axis; the smaller the better) vs. the ``improvement score'' it adds to RIPE RIS (y-axis; the larger the better).

For detailed description and results see our online tool.

Top-5 recommendations for extra RIPE RIS peers and RIPE Atlas probes

A quick view for 5 recommendations for extending the RIPE RIS and Atlas infrastructures with vantage points in ASes that could decrease the bias. For more data, details, and options see the online tools for RIPE RIS and RIPE Atlas

Top-5 recommendations for RIPE RIS
Top-5 recommendations for RIPE Atlas

Subsampling from Internet measurement platforms to decrease measurement bias

One can select subsets ("subsampling") of the vantage points of the Internet Measurement Platforms (IMPs), such as, RIPE Atlas, RIPE RIS, or RouteViews, in order to decrease their bias and get significantly more representative measurement data. In this documentation, we provide (i) detailed results for the bias of subsampling in IMPs, and (ii) a method for efficient subsampling.

The AI4NetMon API

We provide an open API, which provides the bias scores for different IMPs of custom sets of vantage points. The API can be accessed at, and at its endpoints are described in this documentation.

Articles / Blog posts / Presentations

  • Pavlos Sermpezis "Bias in Internet Measurement Platforms" [Blog/Tool/Data], in Observable blog post, 27 October 2022. [ Link]

  • Pavlos Sermpezis "Bias in RIPE RIS per Route Collector" [Blog/Tool/Data], in Observable blog post, 27 October 2022. [ Link]

  • Pavlos Sermpezis "Extending RIPE Atlas in an unbiased way" [Blog/Tool/Data], in Observable blog post, 17 August 2022. [ Link]

  • Pavlos Sermpezis "Extending RIPE RIS in an unbiased way" [Blog/Tool/Data], in Observable blog post, 29 June 2022. [ Link]

  • Pavlos Sermpezis & Alun Davies "Revealing Bias in Internet Measurements" [Podcast], in RIPE Labs, 29 Jun 2022. [ Link]

  • Pavlos Sermpezis "Bias in Internet Measurement Infrastructure" [Presenation], in RIPE84, 17 May 2022. [ Link]

  • Pavlos Sermpezis "Bias in Internet Measurement Infrastructure", in RIPE Labs, 28 March 2022. [ Link]

  • Pavlos Sermpezis "Potential deployments of extra RIPE RIS monitors: Bias vs. Improvement", in Observable blog post, 14 February 2022. [ Link]

  • Pavlos Sermpezis "Estimating the Impact of a Hijack: Measurement Bias and How to Avoid It" [Presenation], in RIPE83 (Routing WG), 22 November 2021. [ Link]

  • Gergana Petrova "RACI Funding in 2021 - Planning and Projects", in RIPE Labs, 17 November 2021. [ Link]



Pavlos Sermpezis, et al. "Unbiasing Internet Measurement Platforms", under submission, 2022.


Dimitrios P. Giakatos, Sofia Kostoglou, Pavlos Sermpezis, Athena Vakali "Benchmarking Graph Neural Networks for Internet Routing Data", in ACM CoNEXT - GNNet workshop, 2022.
[ PDF] [ Code ]


Pavlos Sermpezis, Vasileios Kotronis, Konstantinos Arakadakis, Athena Vakali "Estimating the Impact of BGP Prefix Hijacking", in IFIP Networking conference, 2021.
[ PDF] [ Presentation ] [ Code ]


AI4NetMon [ github repo ]


datalab logo auth logo RIPE logo mpi logo

datalab logo auth logo

RIPE logo mpi logo

Contact (Pavlos Sermpezis)




AI4NetMon is funded by RIPE NCC, through the RACI project funding 2021 program.