arXiv:2610.34635
Quantum Anomaly Detection in network traffic using Variational Quantum Classifiers
Mir H M Ilham, Hardique Dasore, Abdulrahman Ababneh, Hafidz A Kabira, Samyam Shrestha, Jeyasri Sundar, Shah Haque, Eric Howard
Abstract
The increasingly complex cybersecurity landscape requires more advanced network anomaly detection systems. Although, classical machine learning models are known to be effective, they reach computational bottlenecks when processing high-dimensional data. This study investigates the potential of Quantum Machine Learning (QML) for network traffic anomaly detection by evaluating four distinct Variational Quantum Classifier (VQC) architectures. A synthetic network traffic dataset is generated for the evaluations. The VQC models systematically varied feature encoding methods, parameterized ansatz circuits, quantum neural network types, and classical optimizers. Experimental results show that a trade-off exists between circuit expressivity and optimization stability. Deeper circuits (VQC-4) achieved higher accuracy of upto 89.0\%, while simultaneously being sensitive to training conditions. A shallower variant (VQC-1) under performed, although it had consistent and stable optimization curves. Class-dependent biases were also observed such as in VQC-2, where the model struggled to find an effective decision boundary, and produced a large number of false negatives. The best performing model in this study (VQC-3) combined ZFeatureMap with RealAmplitudes ansatz, optimized using ADAM. Comparative analysis against classical machine learning models such as Support Vector Machine and Random Forest reveal a performance gap, where classical models are able to achieve perfect classification performance on the synthetic data. It is concluded that classical algorithms remain dominant in their ability to classify structured data with relatively simple class separations. However, due to the observed potential of certain VQC configurations to learn decision boundaries to classify data, their utility for future applications in noisier, higher-dimensional problems is highlighted.
AI Summary
This paper examines how variational quantum classifier (VQC) configurations affect supervised anomaly detection on synthetic network traffic. It compares four four-qubit architectures that jointly vary feature maps, ansatz depth, estimator- versus sampler-based outputs, and classical optimizers. Experiments use noiseless Qiskit Aer statevector simulation, with sensitivity tests varying initialization, dataset size, class balance, train–test split, and optimization budget. The baseline dataset contains 1,000 samples with four features, balanced labels, and a 70:30 train–test split. Normal and anomalous traffic have distinct statistical distributions; protocol identifiers also separate the classes directly. With 100 optimizer iterations, baseline accuracies are 72.0%, 57.3%, 84.7%, and 89.0% for VQC-1 through VQC-4, respectively. The deepest configuration, using ZZFeatureMap, RealAmplitudes, and COBYLA, obtains the highest baseline accuracy but falls to 78.3% and 76.3% under two alternative seeds. ZFeatureMap with a three-repetition RealAmplitudes ansatz and ADAM shows more consistent performance across the reported sensitivity tests. Classical RBF-SVM and random-forest baselines both achieve 100% accuracy. The reported evidence consists of classification metrics, training-loss curves, confusion matrices, and PCA-based misclassification analysis. The paper identifies the synthetic dataset’s low complexity and strong class separability as limitations of the benchmark. Statevector simulation excludes hardware noise, and the study does not establish performance on real traffic or higher-dimensional tasks. Extensions to temporal data, noise-aware circuits, and automated hyperparameter optimization are left for future work.
Central claim may not hold: The claimed expressivity–optimization-stability trade-off is not isolated by the experiments because circuit depths, encoding choices, QNN types and optimizers vary together, leaving only a comparison of complete training pipelines. The paper incorrectly attributes changes in accuracy to evaluation-set size itself, although a larger test set does not inherently lower expected accuracy and the experiments also change training conditions. The synthetic dataset makes protocol identifier a perfect class discriminator, so its perfect classical classification results establish performance on a trivial constructed task rather than realistic network-anomaly detection.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 28 pages, 23 figures