IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT)
Noise-Aware QML: Adaptive Noise-Regularized Variational Quantum Circuits for Parameter-Efficient Tabular Anomaly Detection
Saksham Garg, Dhruti A, Panchadip Bhattacharjee, Somyajeet Arukh, Arya Abnish Singh, Gururaj H L
Abstract
Anomaly detection on high-dimensional tabular data requires models that are simultaneously compact, robust to noise, and interpretable, which are not met well by traditional ensemble learning approaches under strict parameter budgets. This work introduces NoiseAware-QML, a Variational Quantum Circuit (VQC) approach that incorporates a learnable noise parameter$\sigma$directly into the quantum ansatz, allowing the circuit to adaptively learn noise robustness during optimization through the reparameterization trick. The proposed 4-qubit, 4-layer VQC requires only 54 trainable parameters, a$715 \times$reduction over Random Forest and$91 \times$reduction over XGBoost, while maintaining a multi-seed mean ROC-AUC of$0.8366 \pm 0.0083$(95% CI: [0.8252, 0.8446]) on UNSW-NB15. Three robustness advantages with respect to parameter budget are formally established: (i) robustness to noise-1.84% AUC loss at$\sigma=0.10$compared to 4.12% for Random Forest (2.24× better) and 2.41% for a comparable-sized Dropout-MLP$(1.31 \times$better); (ii) calibration-$\mathbf{E C E}=0.0726$, improving over Random Forest (0.1127) and MLP (0.1151) by 36-37% without calibration correction; (iii) cross-dataset stability-only 5.36% AUC loss on NSL-KDD compared to 8.47% for Random Forest. Ablation study over three seeds verifies that adaptive noise regularization always improves over all variants by$\approx 0.089$AUC. Explainability through quantum-native means, per-gate gradientnorm heatmaps, and Bloch sphere analysis, demonstrates that the second-highest PCA variance component is actually harmful to anomaly detection, a result hidden from classical feature selection. Simulation-based deployment analysis confirms that the model generalizes to simulated NISQ hardware settings with AUC degradation of less than 2%.