Contrastive Structured Detection and Quantum Exploration for Zero-Day Threat Detection

Zero-day threat detection requires models to recognize attack families absent from training data. We investigate a geometry-first approach in which supervised contrastive learning structures a network-flow latent space before anomaly scoring, and we explore a quantum transfor- mation as a post-latent refinement. Saved leave-family-out experiment records show large classical contrastive gains over a reconstruction autoencoder: mean AUC-ROC rises from 0.549 to 0.917 on CICIoT2023 and from 0.791 to 0.995 on UNSW-NB15. We then couple a frozen contrastive encoder to an 8-qubit, 3-layer parameterized quantum circuit (PQC), which completes simulator training at 16K and 32K pilot scales. Latent- space visualizations show that contrastive learning and the PQC produce markedly different representation geometry. The quantum study is inten- tionally exploratory: its current pilot protocol requires a fully leakage-free rerun before any comparative detection or quantum-advantage claim can be made. The contribution is therefore a focused empirical foundation–classical structured detection evidence, a reproducible PQC feasibility demonstration, and a clear basis for controlled quantum evaluation. This is an exploratory preprint and has not been peer reviewed.