Adaptive Hybrid Adversarial Training for Robust Deep Learning-Based Network Intrusion Detection

Abstract: Deep Learning (DL) has become central to modern Network Intrusion Detection Systems (NIDS), yet convolutional and recurrent detectors remain highly susceptible to adversarial perturbations that can silently flip malicious traffic into benign predictions. Existing single-technique adversarial training approaches, most commonly built around the Fast Gradient Sign Method (FGSM) or Projected Gradient Descent (PGD) alone, improve robustness against the specific perturbation family used in training but generalize poorly to unseen attack strategies and often erode clean-data accuracy. This study proposes an Adaptive Hybrid Adversarial Training (AHAT) framework that combines multiple adversarial example generators (FGSM, PGD, and Carlini–Wagner) under an epoch-adaptive curriculum scheduler, fused with a hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) detector and a composite loss that jointly penalizes clean-sample error, adversarial-sample error, and a consistency term between clean and adversarial representations. The framework was evaluated using a quantitative experimental design on benchmark-style network flow data structured after the NSL-KDD and CICIDS2017 traffic schemas, with illustrative/simulated results used to demonstrate the evaluation pipeline pending full empirical deployment. Under this illustrative pipeline, the proposed AHAT model sustained a mean adversarial detection accuracy of 89.1% across FGSM, PGD, and Carlini–Wagner attacks, compared with 82.6% for ensemble adversarial training, 76.9% for standard single-method adversarial training, and 27.5% for an undefended baseline CNN, while clean accuracy was retained at 98.3%. These findings suggest that adaptive, curriculum-driven hybridization of attack families and model architectures offers a more balanced robustness–accuracy trade-off than static, single-technique adversarial training. The paper contributes a reproducible architecture, a composite loss formulation, and an evaluation protocol that other researchers can instantiate on live SPSS-analyzed or network-captured datasets. Implications for security operations centres, cyber-defense policy in developing-country contexts, and future research on certifiably robust NIDS are discussed. Keywords: Adversarial Training; Network Intrusion Detection; Deep Learning; Robustness; Curriculum Learning; CNN-BiLSTM; Cybersecurity. Title: Adaptive Hybrid Adversarial Training for Robust Deep Learning-Based Network Intrusion Detection Author: Samuel Okechukwu Nnaji International Journal of Recent Research in Mathematics Computer Science and Information Technology ISSN 2350-1022 Vol. 13, Issue 1, April 2026 - September 2026 Page No: 56-65 Paper Publications Website: www.paperpublications.org Published Date: 25-August-2026 DOI: https://doi.org/10.5281/zenodo.22094745 Paper Download Link (Source) https://www.paperpublications.org/upload/book/Adaptive%20Hybrid%20Adversarial%20Training-25082026-3.pdf