Evasion in Network Threat Detection Models Using Decision Space Properties

• Rafael Dias Campos UFMG
• Michele Nogueira UFMG / UFPR
• Marcio Costa Santos UFMG

DOI: https://doi.org/10.5753/sbrc.2026.19947

Abstract

Machine learning techniques are frequently employed in detecting network threats, such as denial-of-service attacks, XSS, and Ransomware. However, the trained models can themselves become targets for attacks aimed at evading detection when sending malicious network traffic in the environment and preventing alerts or defensive actions from taking place. Current methods for conducting such attacks present limitations during their training phases and are susceptible to generating traffic samples that differ significantly from the original ones, limiting their utility. To address these issues, this work presents a new method for evading traffic detection by mapping the decision space of the network threat detection models as a set of convex polytopes. The developed method was able to generate samples closer to the original ones when compared with the other methods and presented more stable results among multiple runs.

References

Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M. (2019). Optuna: A nextgeneration hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.

Booij, T. M., Chiscop, I., Meeuwissen, E., Moustafa, N., and Hartog, F. T. H. d. (2022). Ton iot: The role of heterogeneity and the need for standardization of features and attack types in iot network intrusion data sets. IEEE IoT-J, 9(1):485–496.

Ducoffe, M. and Precioso, F. (2018). Adversarial active learning for deep networks: a margin based approach.

Enea (2025). Enea qosmos threat detection sdk - ids-based threat detection. [link]. Acessado em: 14/10/2025.

Fan, F.-L., Huang, W., Zhong, X., Ruan, L., Zeng, T., Xiong, H., and Wang, F. (2024). Deep relu networks have surprisingly simple polytopes.

Fortinet (2025). Next Generation Firewall (NGFW) - See top products. [link]. Acessado em: 21/01/2026.

Gajer, P. and Ravel, J. (2025). The geometry of machine learning models. García, S., Grill, M., Stiborek, J., and Zunino, A. (2014). An empirical comparison of botnet detection methods. Computers & Security, 45:100–123.

Ibitoye, O., Shafiq, O., and Matrawy, A. (2019). Analyzing adversarial attacks against deep learning for intrusion detection in iot networks. In GLOBECOM, pages 1–6.

Koroniotis, N., Moustafa, N., Schiliro, F., Gauravaram, P., and Janicke, H. (2020). A holistic review of cybersecurity and reliability perspectives in smart airports. IEEE Access, 8:209802–209834.

Lin, Z., Shi, Y., and Xue, Z. (2022). IDSGAN: Generative Adversarial Networks for Attack Generation Against Intrusion Detection, page 79–91. Springer International Publishing.

Moustafa, N. (2021). A new distributed architecture for evaluating ai-based security systems at the edge: Network ton iot datasets. Sustainable Cities and Society, 72:102994.

Sattelberg, B., Cavalieri, R., Kirby, M., Peterson, C., and Beveridge, R. (2020). Locally linear attributes of relu neural networks.

Shu, D., Leslie, N. O., Kamhoua, C. A., and Tucker, C. S. (2020). Generative adversarial attacks against intrusion detection systems using active learning. In Proceedings of the 2nd ACM Workshop on Wireless Security and Machine Learning, WiseML ’20, page 1–6.

Tavallaee, M., Bagheri, E., Lu, W., and Ghorbani, A. A. (2009). A detailed analysis of the kdd cup 99 data set. In IEEE CISDA, pages 1–6.

Progress Software Corporation (2025). Flowmon - trusted solution for network and security operations - progress flowmon. [link]. Acessado em: 14/10/2025.