Adversarially Robust Intrusion Detection in Software-Defined Networks Using a Resilient Adversarial Deep Neural Network with Stochastic Gradient Langevin Dynamics

Software-Defined Networking (SDN) has transformed network management by separating the control plane from the data plane, thereby improving flexibility, programmability, and centralized control. However, the centralized architecture also introduces security challenges, making SDN environments attractive targets for cyberattacks. Although deep learning-based Intrusion Detection Systems (IDSs) have demonstrated high detection performance, they remain vulnerable to adversarial attacks that can drastically degrade classification accuracy. This study presents an adversarially robust intrusion detection system for SDN by incorporating a Resilient Adversarial Deep Neural Network (RANetDNN) with Stochastic Gradient Langevin Dynamics (SGLD) optimization approach. The model was implemented and evaluated using the InSDN dataset and compared with a Deep Neural Network (DNN) baseline. Performance was evaluated using Accuracy, Precision, Recall, F1-score, ROC-AUC and adversarial robustness under Fast Gradient Sign Method (FGSM) attacks. Experimental results show that while the baseline DNN experienced a drastic reduction in F1-score under increasing adversarial perturbations, the RANetDNN-SGLD model consistently maintained high detection performance and robustness across all evaluated perturbation levels. The experimental results demonstrate that the RANetDNN-SGLD model achieved excellent intrusion detection performance while showing significantly greater robustness against FGSM adversarial attacks than the baseline Deep Neural Network. Comparative evaluation showed that this approach maintained consistently high F1-score values under increasing adversarial perturbations, highlighting its effectiveness for securing Software-Defined Network environments against both conventional cyber threats and adversarial attacks.

Keywords: Software-Defined Networking (SDN), Intrusion Detection System (IDS), Resilient Adversarial Network (RANet), Stochastic Gradient Langevin Dynamics (SGLD), Deep Learning, Deep Neural Network.