By Enrique García Ceja Jesús Arturo Pérez Díaz
The digital transformation has connected millions of people and businesses across Latin America, but it has also expanded the attack surface available to cybercriminals.
Today, a cyberattack can cripple hospitals, disrupt public services, compromise financial institutions, or bring Internet service providers to a standstill.
The research study “3MuViS, 3-Phase Multi-View Stacking: A Model Selection Algorithm for Multi-View Fusion” aims to develop artificial intelligence models capable of detecting threats across critical Internet infrastructure, including Internet service providers (ISPs), Internet exchange points (IXPs), transit providers, and Internet of Things (IoT) devices.
What Is Multi-View Learning?
Today’s cyberattacks generate massive amounts of data. Every network connection produces different types of information, including traffic characteristics, communication statistics, and device behavior. Traditionally, many detection systems analyze only a subset of these data, making it harder to identify increasingly sophisticated attacks.
Overcoming this limitation requires an approach known as Multi-View Learning. Rather than analyzing a single representation of the data, these models simultaneously examine multiple “views” of the same event, each providing complementary information about network behavior.
An artificial intelligence model then combines these perspectives to reach a more accurate decision than any of the individual models could achieve on its own.
This strategy, known as Multi-View Stacking, combines the strengths of different algorithms to improve the detection of complex threats.
Three-Phase Multi-View Stacking
As part of the project, researchers at Tecnológico de Monterrey developed 3MuViS (Three-Phase Multi-View Stacking), an algorithm that automates the selection of the most suitable AI models for each data source.
The results show that, rather than relying on an extensive trial-and-error process, the system automatically identifies the best combination of algorithms for each task, significantly reducing computational costs without compromising accuracy.
Evaluations using cybersecurity datasets and data from other applications showed that this approach consistently outperforms traditional strategies. It also delivers results close to those of the best possible model configurations while requiring only a fraction of the processing time.
These findings highlight the potential of Multi-View Learning to build more robust and adaptable detection systems. The methodology also extends beyond cybersecurity to fields such as healthcare, human activity recognition, and biomedical analysis.
An Open-Source Software Library
Another key outcome of the project was the development of multiviewstacking, an open-source software library for the Python programming language. It allows researchers and developers to build multi-view models without having to implement the algorithms from scratch [2].
The library makes it easier for other research teams to reproduce, validate, and build on these advances, accelerating innovation across multiple disciplines.
This research is particularly relevant to Latin America. According to recent reports from specialized cybersecurity companies, the region faces hundreds of billions of cyberattack attempts every year.
Among the most common threats are ransomware, which locks users out of their systems or encrypts their data; distributed denial-of-service (DDoS) attacks, which overwhelm a system or server with requests, preventing legitimate users from accessing it; credential theft, involving the illicit acquisition of passwords, usernames, and other information to gain access to accounts or systems; and malware, or malicious software. These threats affect both public- and private-sector organizations.
At the regional level, Brazil, Mexico, and Colombia accounted for nine out of every 10 recorded attacks in Latin America [3].
For countries undergoing rapid digitalization, detecting an attack in its early stages can prevent multimillion-dollar financial losses, protect essential services, and limit the impact on millions of users. AI models capable of analyzing massive volumes of data in real time offer a strategic advantage in strengthening the region’s digital resilience.
References
- Villegas-Morales, G., Garcia-Ceja, E., Hinojosa, S., Pérez-Díaz, J. A., & Zareei, M. (2026). 3MuViS, 3-Phase Multi-View Stacking: A Model Selection Algorithm for Multi-View Fusion. Machine Learning and Knowledge Extraction, 8(8), 227.
- Garcia-Ceja, E. (2026). multiviewstacking: A Python Package for Training Multi-View Stacking Classifiers. Journal of Open Research Software, 14(1), 41.
- Statista. (2025) Cyber crime in Latin America – statistics & facts. Updated Dec 17, 2025.
Authors
Enrique García Ceja. Professor in the Department of Computing at Tecnológico de Monterrey, Mexico (2022–present). He previously worked as a data scientist at Optimeering and as a researcher at SINTEF, both in Norway. He completed a postdoctoral fellowship at the University of Oslo and earned his Ph.D. and master’s degree in Intelligent Systems from Tecnológico de Monterrey, Mexico. He has also worked as a software engineer at Huawei and is the author of Behavior Analysis with Machine Learning Using R.
Jesús Arturo Pérez Díaz. Ph.D. in Computer Science from the University of Oviedo. He is a professor at Tecnológico de Monterrey’s Guadalajara campus and a Level I member of Mexico’s National System of Researchers (SNII). His cybersecurity research focuses on detecting and mitigating cyberattacks, including high-rate and low-rate distributed denial-of-service (DDoS) attacks and ransomware, using AI and generative AI models. He has also designed and deployed frameworks for software-defined networking (SDN).

