Benchmarking the adversarial resilience of machine learning models for DDoS detection

Distributed Denial of Service (DDoS) attacks continue to grow in scale and sophistication, making timely and reliable detection increasingly challenging. Machine learning (ML) models have demonstrated promise in identifying malicious traffic patterns. However, their vulnerability to adversarial manipulation remains a critical security concern. This study benchmarks the adversarial robustness of several standalone ML models trained on clean traffic from the CIC-IDS2017 dataset. Adversarial perturbations are generated using the Fast Gradient Sign Method (FGSM). Baseline robustness is assessed by evaluating each model on FGSM adversarial samples generated at ϵ = 0 . 1 . Adversarial training is then performed by augmenting the clean dataset with FGSM-generated samples, after which models are evaluated across multiple perturbation strengths ( ϵ = 0 . 1 to 3 . 0 ). Experimental findings show that boosting-based models, particularly XGBoost and LightGBM, demonstrate the highest resilience under adversarial stress. In contrast, models such as Logistic Regression and MLP experience significant performance degradation, even after adversarial training. Despite this, adversarial training substantially improves robustness across all models, highlighting its effectiveness in mitigating FGSM-induced decision boundary shifts. This benchmarking study underscores the importance of adversarial training and model selection when deploying ML-based intrusion detection systems. Boosting ensembles consistently provide superior robustness, while linear and neural models remain more susceptible to perturbations.