Toward Adversarial Robustness Network Intrusion Detection Based on Multi-Model Ensemble Approach

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Machine learning-based network intrusion detection systems (NIDS) are susceptible to adversarial attacks.
• Existing robustness research for tabular NIDS data is limited by single-model, single-dataset, and non-adaptive evaluations.

Purpose of the Study:

• To conduct a comparative robustness study of a four-component defense pipeline for tabular NIDS.
• To evaluate the dataset and architecture dependence of adversarial defenses.

Main Methods:

• Evaluated XGBoost, LightGBM, TabNet, and Residual MLP on RT_IOT2022 and Web_IDS23 datasets.
• Assessed performance under standard, constrained, and adaptive attacks, including component-wise ablations and sensitivity analyses.
• Measured per-class F1 scores and computational overhead.

Main Results:

• Defense effectiveness shows strong dependence on dataset and model architecture.
• Tree-based models on RT_IOT2022 reduced robustness gaps but impacted clean accuracy; Residual MLP offered a better balance.
• On Web_IDS23, simpler defenses sometimes outperformed the full pipeline; median filtering proved to be the most fragile component.

Conclusions:

• Adversarial defense for tabular NIDS is validation-driven and dataset-specific.
• The evaluated four-component defense stack is not a universal solution and requires careful tuning.
• Deployment limitations include minority-class collapse and significant training costs.