Adversarial-Trained ensembles for robust intrusion detection
|
Full Text |
Pdf
|
|
Author |
Ummadisetty Venugopal and Mediga Nagaseshudu
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
606-616
|
|
Volume No. |
21
|
|
Issue No. |
9
|
|
Issue Date |
July 10, 2026
|
|
DOI |
https://doi.org/10.59018/052667
|
|
Keywords |
intrusion detection system (IDS), adversarial machine learning, adversarial training, deep learning, ensemble learning, DDoS attacks, network security.
|
Abstract
The paper will explore the application of adversarially trained deep ensembles to robust network intrusion detection. Although missing benign conditions modern ML-based intrusion detection systems (IDSs) can achieve high accuracy, adversarial perturbations that misclassify can be created and induce a lack of trust in security operations. To solve the issue, we will create an adversarially trained ensemble model, which consists of a 1-D CNN-LSTM classifier with adversarial example generation via gradients in the training loop. The framework is tested on the CICIDS2017 DDoS traffic bit subset, with default pre-processes, and stratified train test partition. The experimental findings indicate that the proposed model has 99.65 percent accuracy, 99.47 percent precision, 99.92 percent recall, 99.70 percent F1-score, and ROC-AUC of 0.9999 on clean test data and maintains 99.96 percent accuracy given FGSM perturbation in a bounded threat model. These results point to the fact that the ensemble trains graphical representations that are perturbation-invariant without any loss in detection. Moreover, a real-time DDoS threat simulator is created to reveal the model in an interactive interface to enable the operators to manipulate the principal traffic parameters and receive the resulting threat scores, which also improves the interpretability and practical usability. In general, this paper has shown that adversarially trained ensembles are able to substantially increase the resilience of IDSs against newly developed adversarial machine learning threats.
Back