Machine learning-based network intrusion detection systems (NIDS) require representative training data to detect evolving cyber-attacks. In order to identify specific attack types, multi-class models promise to be a suitable approach, but imbalanced training datasets make it difficult to train multi-class models to reliably detect specific cyber-attack types. Although prior work claims that resampling improves the classification performance, the literature is missing a comprehensive evaluation of classical and generative resampling methods, alongside a detailed assessment of synthetic data quality.
Our study fills that gap by assessing how different resampling strategies affect synthetic data quality and classifier performance. We evaluate 42 resampling combinations on seven classification models trained on five datasets and show differences in the quality of the synthetic data. Interestingly, synthetic data that is less similar to the original data often improve the classification results.
| Titel | Rethinking Synthetic Oversampling for Intrusion Detection: When Similarity Hurts Performance |
|---|---|
| Medien | Computational Intelligence in Security for Information Systems |
| Verlag | Springer |
| Herausgeber | Emilio Corchado, Héctor Quintián, Hilde Pérez García, José Luis Calvo Rolle, Sérgio Filipe Ramos, Francisco Javier Martínez de Pisón, Paolo Fosci |
| Band | 3013 |
| Verfasser | Maximilian Wolf, Prof. Dr. Dieter Landes, Andreas Hotho, Daniel Schlör |
| Seiten | 16-27 |
| Veröffentlichungsdatum | 18.06.2026 |
| Projekttitel | BRACE-LLM |
| Zitation | Wolf, Maximilian; Landes, Dieter; Hotho, Andreas; Schlör, Daniel (2026): Rethinking Synthetic Oversampling for Intrusion Detection: When Similarity Hurts Performance. Computational Intelligence in Security for Information Systems 3013, 16-27. |