Implications of Applying Emulation Data to Machine Learning-Based Intrusion Detection Systems

Abstract

Intrusion detection systems (IDS) require high-quality benchmark data sets for evaluating their performance in distinguishing suspicious activities from normal behavior patterns. While real-world network environments often consist of physical hardware components, it is, however, difficult to generate an IDS benchmark data set in a real-world network environment for reasons such as scalability, costs, anonymity in real-world data, infrastructure, and availability of hardware components. A solution to this problem is the usage of emulated network environments to create IDS benchmark data sets. Nevertheless, it is unclear whether emulated data represents physical network characteristics based on the recorded data patterns. Several IDS benchmark data sets were generated in the past, but none addressed the described infrastructure gap so far. To address this issue, we recreated a small physical enterprise network by using an emulated network environment. Inside both networks, we executed identical script-based user behavior, which includes six normal behavior categories and one attack behavior category. To check whether the emulated data are a suitable proxy for training machine learning-based IDS, recorded host and network data sources of both simulation environments are analyzed via diverse approaches. We use Wasserstein distances to compare the properties of the data distributions. Subsequently, we apply established classification metrics to evaluate the performances of a collection of machine learning-based classifiers on the data. For both host and network data, differences between emulated and actual hardware-based data could be identified. In general, the machine learning-based performances show that emulated IDS data is a good proxy for hardware-generated data produced on actual devices during our experiments.

Mehr zum Titel

Titel Implications of Applying Emulation Data to Machine Learning-Based Intrusion Detection Systems
Medien Journal of Information Security and Applications (JISA)
Verlag Elsevier
Band 97
Verfasser Maximilian Wolf, Kevin Bergner, Pascal Förtsch, Prof. Dr. Dieter Landes
Seiten 104362
Veröffentlichungsdatum 17.01.2026
Zitation Wolf, Maximilian; Bergner, Kevin; Förtsch, Pascal; Landes, Dieter (2026): Implications of Applying Emulation Data to Machine Learning-Based Intrusion Detection Systems. Journal of Information Security and Applications (JISA) 97, 104362.