The growing sophistication of cyber threats and adversarial attacks poses critical challenges to the security and robustness of machine learning models deployed in real-world systems. While traditional deep learning architectures excel in clean data classification, they often fail under adversarial perturbations, exposing vulnerabilities in sensitive domains such as healthcare, finance, and industrial control. In this paper, we introduce a novel hybrid adversarially-trained deep learning framework that integrates reinforcement learning-inspired robustness adaptation with knowledge-driven regularization to achieve improved resilience against fast gradient sign method (FGSM) and projected gradient descent (PGD) attacks. Our approach simultaneously optimizes clean accuracy and adversarial robustness by balancing cross-entropy and adversarial loss components, while monitoring calibration error, gradient dynamics, and generalization gap to ensure stable convergence. Extensive experiments on reconnaissance, shellcode, and worms datasets demonstrate that the proposed model achieves up to 97.88% accuracy on clean data and maintains 84.9% accuracy under FGSM and 81.75% under PGD attacks, outperforming convolutional neural network (CNN) and long short-term memory (LSTM) baselines by more than 6–10 percentage points in adversarial robustness. Furthermore, the training curves reveal consistent improvements in convergence stability, runtime efficiency, gradient norm decay, and a 30% reduction in expected calibration error, validating the scalability of the framework. This work contributes not only a novel adversarial defense paradigm but also provides insights into the trade-offs between robustness, efficiency, and generalization.
Saha et al. (Mon,) studied this question.