AN ADVERSARIALLY ROBUST RSHO DBIGRNN FRAMEWORK FOR DDOS INTRUSION DETECTION

Abstract

DDoS-Distributed Denial-of-Service intrudes constitute an important threat to network obtainability by exhausting network and computational sources via harmful traffic. Though machine learning(ml) and deep learning based Network Intrusion Detection Systems (NIDSs) can determine complex traffic design, high identification on clean data does not essentially imply robustness towards adversarial manipulation. This research implements an adversarially robust extension of the Refined Selfish Herd Optimization oriented Deep Bidirectional Gated Recurrent Neural Network (RSHO-DBiGRNN) prototype for DDoS intrusion detection. The prototype stays the core feature illustration, Hybrid Vulture–Dragonfly Optimization Algorithm (HVD-OA) feature selection, and RSHO-DBiGRNN elements while incorporating leakage-controlled, training-fitted preprocessing, multi-attack adversarial training, and robustness-aware configuration optimization. Network traffic dataset feature extracted using parallel Autoencoder (AE), Linear Discriminant Analysis (LDA), and Adaptive Entropy-Weighted Non-Negative Matrix Factorization (NMF) transition, succeeded by HVD-OA oriented feature selection. Adversarial samples are created using the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and DeepFool in the chosen feature space and incorporated into model training alongside clean samples. Moreover, the RSHO optimization objective is reformulated to select recurrent dropout configurations utilizing a connected clean and multi-attack adversarial validation criterion. Execution is evaluated under clean and adversarial states making use of accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and receiver operating characteristic area under the curve (ROC-AUC). Threat specific execution degradation is also measured relative to clean test performance. The research looks into the contribution of combining multi-attack adversarial training with robustness-aware configuration chosen to DDoS finding resilience subjected to evaluated feature-space perturbations. Experimental output displays that the implemented model achieves 99.24% accuracy and 99.37% F1-score under clean test conditions. Under FGSM, PGD, and DeepFool attacks, it achieves F1-scores of 97.03%, 96.15%, and 96.97%, correspondingly.