Journal of Applied Science and Engineering

 Copyright The Author(s). This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are cited.

Download Citation:  BibTeX | http://dx.doi.org/10.6180/jase.202610_33.011  

Download PDF

With the rapid digitization of critical infrastructure and the proliferation of complex cyber threats, deep learning (DL) has become a core technology for network threat classification, enabling accurate identification of malicious activities such as DDoS attacks, malware infections, and phishing attempts. However, DL models are inherently vulnerable to adversarial attacks, subtle, human-imperceptible perturbations added to input network traffic features that can misleadingly alter model predictions, posing severe risks to network security. Additionally, the black-box nature of most advanced DL models (e.g., Transformers, deep neural networks) hinders their practical deployment in security-critical scenarios, as security analysts cannot interpret the
reasoning behind classification decisions. To address these two critical challenges with robustness against adversarial attacks and model explainability, this paper proposes a novel Robust and Explainable Deep Learning Framework (REDLF) for network threat classification. The framework integrates three core components: (1) an Adversarial Training Module (ATM) based on projected gradient descent (PGD) and physical environment modeling to enhance model robustness; (2) a Feature Enhancement Module (FEM) that combines attention mechanisms and variational autoencoders (VAEs) to extract discriminative and robust traffic features; (3) an Explainable Interpretation Module (EIM) that fuses SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to generate both global and local explanations for classification decisions. Theoretical analysis proves the convergence and robustness of the proposed framework, and extensive experiments are conducted on three benchmark network threat datasets (CSE-CIC-IDS2018, NSL-KDD, and CTU-13) under six typical adversarial attacks (FGSM, PGD, C&W, JSMA, BIM, and DeepFool). Experimental
results demonstrate that REDLF outperforms state-of-the-art methods in both classification accuracy and adversarial robustness achieving an average accuracy of 98.72% on clean data and maintaining an accuracy of over 92.35%understrongadversarial attacks, which is 7.89% higher than SOTA models on average. Furthermore, the EIM module provides intuitive, human-interpretable explanations that clarify the key traffic features (e.g., flow duration, packet length, and protocol type) influencing classification decisions, addressing the black-box problem of DL models. This work contributes to bridging the gap between robustness and explainability in DL-based network threat classification, providing a reliable and interpretable solution for real-world network security defense.

Keywords: Deep Learning; Network Threat Classification; Adversarial Attacks; SHAP; Adversarial Training