QUANTUM-IDS: a quantum-inspired deep learning framework with multi-scale temporal analysis and meta-learning for network intrusion detection
Network intrusion detection systems currently face key challenges in detecting advanced cyber-attacks that have non-linear feature relationships, various scales of temporal relationships, and enhanced evasion techniques. This study investigates the applicability and empirical value of five concepts for network intrusion detection by embedding them in a deep learning architecture called QUANTUM-IDS: Quantum Feature Entanglement (QFE), a classical feature-mixing mechanism that uses feature expansion and repeated learned pairwise interactions; a Multi-Scale Temporal Convolutional Network (MS-TCN) for representing traffic at multiple temporal resolutions; an Adversarial Robustness Module (ARM) combining adversarial training with certified robustness guarantees; a Meta-Learning for Few-Shot Detection module (MFD) intended to support efficient few-shot adaptation within the evaluated attack-category distribution using a handful of labelled examples and a Causal Inference Attribution module (CIA) intended to support explainability of the decision process. Evaluated under a leakage-controlled protocol on a five-class subset of CSE-CIC-IDS2018 (248,347 de-duplicated flows; Benign, DoS, DDoS, Infiltration, Bot; Timestamp excluded), QUANTUM-IDS reached 85.76% ± 1.31% test accuracy and 73.02% macro-F1 over five seeds, below an MLP (88.12%) and gradient-boosted trees (90.62%). Without adversarial training (clean-accuracy variant), it reached 88.83% ± 1.07%. With adversarial training, it kept accuracy at 81.8% under PGD-20 attacks (ε = 0.1) and 81.5% under AutoAttack, against 62.8% and 49.0% for an undefended MLP. Ablations over five paired seeds showed no significant contribution of the quantum-inspired, multi-scale, few-shot or causal modules, and an equally robust adversarially trained MLP indicates that robustness stems from the training procedure. The principal limitation is that attack classes absent from training were not detected reliably, even with labelled examples. A leakage audit showed that including the Timestamp field would raise accuracy from 85.0 to 94.9%.
Authors
- Parthasaradhi Mayasala
- Vishnu Kumar Mishra (ORCID: https://orcid.org/0000-0001-8652-9076)
- S. Satya Narayana
- B. Hari Krishna
- K. Sudhakar
Institutions
- Malla Reddy University (IN)
- Malla Reddy Engineering College for Women (IN)
- Malla Reddy Vishwavidyapeeth (IN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1038/s41598-026-80908-0
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00