Hybrid quantum–classical neural kernel emulator for scalable drive-by attack detection

Drive-by attacks exploit browser vulnerabilities to deliver malicious payloads without user interaction, posing severe threats to web security. Quantum Machine Learning (QML) offers enhanced feature representation through quantum kernels, but faces scalability and hardware accessibility challenges. We propose a hybrid quantum–classical neural kernel emulator that approximates quantum kernel behavior using classical neural networks, enabling scalable drive-by attack detection without quantum hardware at inference. Our framework combines Nyström approximation with k-means landmark selection to construct a low-rank quantum kernel matrix, and employs a tokenized transformer encoder to emulate quantum feature maps from full 128-dimensional HTTP request features. A quantum-anchor subspace of 8 features, selected via mutual information and recursive feature elimination, guides the quantum kernel computation during training, while the neural emulator incorporates all 128 features through structured tokenization. The emulated features feed a classical support vector machine for end-to-end attack detection. Experiments on a DriveBy-HTTP-2026 dataset (45,280 samples) demonstrate 98.3% accuracy and 0.983 AUC while reducing inference latency from 1842 ms for direct quantum-kernel inference to approximately 1.8 ms per request using fully classical deployment. We validate our method on public benchmarks (CICIDS-2017: F1 = 0.9032, AUC = 0.9657; UNSW-NB15: F1 = 0.8462, AUC = 0.8759). Additional ablation, classical-kernel-emulation, sensitivity, and robustness analyses are included to clarify the role of quantum-kernel guidance. Rather than claiming asymptotic quantum advantage, the proposed framework uses quantum-kernel-induced geometry as a training-time structural guide and transfers this guidance into a scalable neural emulator operating on the complete 128-dimensional feature representation.