Hybrid Quantum-Classical Intrusion Detection under Severe Class Imbalance: A Controlled Comparison with a Classical Counterpart

Abstract Intrusion Detection Systems (IDSs) must detect known threats and zero-day attacks in traffic with strong class imbalance. This study evaluates, under controlled conditions, whether a variational quantum circuit (VQC) contributes to classification in a hybrid IDS. Tabular traffic features from CICIoT2023 (five classes) and UNSW-NB15 (binary) are converted to grayscale images and processed by a ResNet18 backbone followed by a dressed head with an eight-dimensional bottleneck. The core of the head is either a VQC with 32 parameters (ResNet18-VQC), simulated without noise, or a classical block with 72 parameters (ResNet18-CDN); everything else is identical. The backbone is reused without external pre-trained weights, either frozen or with Layer 4 fine-tuned together with a genetic algorithm on the output layer, and XGBoost serves as a tabular reference. Across ten seeds, the circuit produced no change in macro F1-score detectable against training variability in either dataset or regime (Holm-adjusted paired t -test, $$p\ge 0.907$$ ). A bootstrap on the test set found a small advantage of the VQC only in the frozen regime of CICIoT2023 (+0.007). XGBoost obtained the highest macro F1-score in both datasets (0.885 and 0.918). The results come from noiseless simulation and do not indicate quantum advantage.