Deep Learning for DDoS Detection: A Systematic Review of Explainability and Adversarial Robustness

Abstract: Deep learning (DL) is outperforming all other methods in Distributed Denial of Service (DDoS) detection; however, there are two important remaining gaps: (i) the interpretability of model decisions is very limited, and (ii) models are susceptible to adversarial attacks that are very subtle traffic manipulations to avoid detection. This systematic literature review (SLR) reviews the research conducted from 2020 until 2025 regarding deep learning-based DDoS detection with particular attention on Explainable AI (XAI) and adversarial robustness. The process follows PRISMA-inspired guidelines whereby searches are conducted on IEEE Xplore, SpringerLink, Web of Science, and associated indexes with the predefined keywords, and clear inclusion/exclusion criteria are applied, focusing on: DL-based DDoS/IDS models, integration of XAI techniques, and/or evaluation or improvement of robustness against adversarial or evasive attacks. The final corpus includes the most recent XAI-enabled DDoS defenses (e.g., SHAP-based interpretable CNN/LSTM, Kolmogorov–Arnold (K–A) networks with XAI, attention-based temporal explainers), adversarial robust frameworks (e.g., GAN-based adversarial training, AutoML-based robust ensembles, CTGAN-based augmentation), and cross-cutting surveys on XAI and adversarial machine learning for IDS. Our synthesis (i) proposes a taxonomy linking model architecture, deployment context, XAI mechanism, and robustness strategy; (ii) demonstrates how XAI improves operator trust, feature selection, and debugging; (iii) shows that adversarial training and realistic perturbation modeling significantly hinder evasion success; and (iv) identifies open challenges for jointly optimizing accuracy, interpretability, and robustness under realistic, evolving network environments. The review ends with a set of research avenues for future DDoS detection systems that will be accurate, typically transparent, and robust.