Swathi Chandrasekhar and Shiva Raj Pokhrel of Deakin University investigated quantum machine learning (QML) for detecting network intrusions, addressing limitations in classical machine learning’s ability to scale and generalize to new threats. Their work presents a comprehensive evaluation of four quantum classifiers, Quantum Support Vector Classification, Quantum Neural Networks and Variational Quantum Classifiers, using the widely adopted BoT-IoT dataset as a benchmark for IoT security.

The researchers aimed to delineate conditions for quantum advantage in anomaly detection, noting that classical machine learning often struggles with “scalability, robustness in high-dimensional spaces, and generalization to novel threats.” This systematic study offers new insights into the expressivity and scalability of quantum-enhanced security methods.

QML Framework for BoT-IoT Anomaly Detection

The study systematically assessed the influence of dimensionality reduction techniques, Behavioral and Statistical Analysis (BSA), PCA, and t-distributed stochastic neighbor embedding (t-SNE), on the effectiveness of quantum classifiers. The framework used to evaluate these algorithms incorporates a two-stage process beginning with feature selection and dimensionality reduction. Feature importance was quantified using behavioral and statistical metrics, including point-biserial correlation, chi-square tests and ANOVA F-tests, with aggregated scores determining the top features for analysis.

This preprocessing step is critical, as the BoT-IoT dataset contains a high number of features that can hinder the performance of both classical and quantum machine learning models. QSVC, for example, offered strong precision with a lower computational burden, suggesting it may be suitable for resource-constrained IoT devices.

Training loss visualizations indicated that VQC converged faster and more stably across dimensionality reduction methods, hinting at its potential for scalable IoT anomaly detection. The researchers state, “The researchers provide a comparative performance evaluation of these quantum approaches against classical baselines, identifying the conditions under which QML models yield advantages, and analyze their scalability, generalization and deployment feasibility for real-world IoT security applications.”

Dimensionality Reduction with PCA and t-SNE

PCA projects data onto a lower-dimensional space by solving an eigenvalue problem, while t-SNE focuses on preserving local similarities through minimization of the Kullback-Leibler divergence between high- and low-dimensional probability distributions. The choice of technique demonstrably impacted the performance of each QML model tested. QSVC, or Quantum Support Vector Classification, delivered strong precision but at a higher computational cost when using PCA, suggesting a trade-off between speed and accuracy.

Conversely, the same QSVC implementation with t-SNE offered efficient convergence alongside strong recall, potentially making it suitable for IoT devices with limited processing power. QNN, or Quantum Neural Networks, exhibited more variable performance across the different dimensionality reduction methods, with lower scores overall compared to VQC and QSVC. Further insight came from visualizations of training loss distributions. VQC with BSA consistently exhibited the narrowest, lowest-centered distribution, confirming both stable training and superior predictive performance.

QSVC on PCA, despite achieving high accuracy, showed the widest spread in loss values, indicating substantial variability across training iterations. The use of t-SNE generally yielded more balanced loss distributions for all algorithms, with QSVC maintaining a consistently lower spread than both QNN and VQC.

“VQC combined with BSA achieved the best overall trade-off between accuracy, stability, and runtime,” the paper reports, highlighting the importance of careful algorithm and technique selection for optimal performance in anomaly detection. The study’s findings suggest that while VQC with BSA currently offers the best overall performance, QSVC with t-SNE presents a viable alternative for resource-constrained environments, demonstrating the potential for quantum machine learning to enhance IoT security.

Quantum Support Vector Classification Implementation

Quantum Support Vector Classification (QSVC) uses quantum kernels to map data into a high-dimensional Hilbert space, a technique designed to capture complex correlations within high-dimensional Internet of Things (IoT) traffic data. The kernel function, expressed as K(x,x’;θ) =, uses parameterized quantum states to define relationships between input vectors, potentially exceeding the capabilities of classical kernel methods.

This approach, detailed in Algorithm 1, begins with dimensionality reduction and feature normalization before optimizing the kernel parameters to minimize support vector machine loss. The study’s results, presented in Table I, indicate that QSVC paired with Behavioral and Statistical Analysis (BSA) achieved an accuracy of 0.85, alongside a precision and F1-score of 0.72 and 0.85 respectively.

Variational Quantum Classifier Training Process

The Variational Quantum Classifier (VQC) training process, detailed in the paper, begins with dimensionality reduction applied to input data before feature normalization. This pre-processing step is integral to both QSVC and VQC algorithms, as outlined in Algorithms 1 and 2 respectively, and prepares the data for optimal quantum processing. The algorithms then initialize parameters before iteratively optimizing them to minimize a specified loss function, a process central to achieving accurate anomaly detection.

Performance metrics varied significantly depending on the dimensionality reduction technique employed. This suggests BSA may be particularly effective at feature engineering for this specific quantum approach. This work proposes a unified QML framework for anomaly detection, instantiated and evaluated on the BoT-IoT benchmark dataset, offering the first systematic study of quantum enhanced methods in this security critical domain.

Quantum Neural Network Output Expectation Value

The QNN output, specifically, is calculated as the expectation value: ⟨, where U(x,θ) represents the parameterized quantum circuit and O is the measurement observable. This formulation allows for the approximation of complex functions, mirroring the functionality of classical neural networks but within a quantum computational framework. Performance evaluations, as detailed in Table I, demonstrate the impact of dimensionality reduction techniques on the efficacy of these quantum classifiers.

Conversely, the same VQC implementation, when used with t-distributed Stochastic Neighbor Embedding, yielded significantly lower results, with an accuracy of 0.74, precision of 0.73 and F1-score of 0.74. The implementation and benchmarking of these four quantum machine learning models, Quantum Support Vector Classifier, Variational Quantum Classifier and Quantum Neural Networks, were conducted using the BoT-IoT dataset. This dataset is a widely adopted benchmark for IoT security, indicating a focus on practical application against a known threat landscape rather than purely theoretical exploration.

👉 More information
🗞 Evaluation of Quantum Machine Learning Models for Network Anomaly Detection
✍️ Swathi Chandrasekhar and Shiva Raj Pokhrel
🧠 ArXiv: https://arxiv.org/abs/2611.09253

More like this