Quantum-Enhanced Anomaly Detection

Updated 23 January 2026

Quantum-enhanced anomaly detection is a method that leverages quantum properties like superposition and entanglement to improve feature representation and efficiency.
The integration of QUBO models, hybrid quantum neural networks, and quantum kernel methods has achieved state-of-the-art metrics, including ROC-AUCs near 0.97 and enhanced F1-scores.
Emerging architectures such as quantum autoencoders, GANs, and federated quantum learning address scalability and noise challenges, enabling robust detection in domains like cybersecurity and finance.

Quantum-enhanced anomaly detection refers to the integration of quantum computing techniques into machine learning models for discovering outliers or anomalous data in complex datasets. This paradigm leverages the quantum properties of superposition, entanglement, and Hilbert space embedding to realize more expressive feature representations, potentially improved generalization, and resource-efficient computations compared to classical approaches. It encompasses a spectrum of methods—ranging from quantum-inspired formulations of classical objectives to hybrid architectures combining variational quantum circuits with classical neural networks—that target tabular, time-series, image, multi-modal, and sequential data in domains such as cybersecurity, industrial monitoring, finance, biomanufacturing, and critical infrastructure.

1. Quantum-Inspired Formulations and QUBO Models

The QUBO (Quadratic Unconstrained Binary Optimization) approach provides a unifying framework for recasting distance- or statistics-based anomaly detection as a binary quadratic optimization problem. In this model, a binary vector $x \in \{0,1\}^N$ encodes anomaly assignments (with $x_i=1$ marking sample $i$ as anomalous). The objective function is constructed as

$\min_{x \in \{0,1\}^N} x^{T} Q x,$

where $Q$ incorporates both statistical (individual deviation from the data centroid) and pairwise (mutual distance among chosen anomalies) contributions: $Q_{ii} = -\alpha d_i + A(1-2k), \quad Q_{ij} = -(1-\alpha) d_{i,j} + 2A \quad (i \neq j).$ Here, $d_i$ denotes a point-to-centroid distance, $d_{i,j}$ is the pairwise distance, $\alpha$ balances the two terms, $A$ is a large penalty parameter, and $k$ enforces cardinality through a quadratic penalty. Quantum and quantum-inspired annealing machines (e.g., using D-Wave QPU architectures) solve these QUBOs by minor-embedding $Q$ onto the device topology and searching for minimum-energy states. Empirical benchmarks demonstrate that the QUBO-based detector achieves state-of-the-art ROC-AUCs (e.g., $\sim$ 0.97 for credit card fraud), matches or exceeds classical methods like EllipticEnvelope, Isolation Forest, and One-Class SVM, and shows enhanced robustness to hyperparameter choices (Mellaerts, 2023).

2. Quantum Neural Networks and Hybrid Deep Learning

Hybrid quantum-classical neural architectures integrate parameterized quantum circuits as either feature mappers or classifier/regressor modules within larger classical learning pipelines. The H-FQNN (Hybrid Fully-Connected Quantum Neural Network) exemplifies this design: an initial dense classical layer feeds a quantum variational circuit (VQC) whose qubit count and ansatz depth are tuned for maximal expressivity and efficient quantum advantage. Features are encoded via angle (rotation) gates, and subsequent variational layers (entangling CNOT patterns) allow the VQC to learn expressive nonlinear decision boundaries. Readout occurs via Pauli- $Z$ measurements, returning expectation values as features for a final dense (classical) head.

Gradient updates in such hybrid models use the parameter-shift rule for quantum circuit parameters, and standard backpropagation for classical weights. On ADS-B dataset benchmarks, an H-FQNN with six qubits slightly outperforms classical FNN baselines using BCEWithLogitsLoss (F1 = 93.29% vs. 93.00%), with strong performance persisting up to moderate quantum layer width (Naaman et al., 19 Sep 2025). Advantages arise from superposition (rich $2^{n_q}$ -dimensional state spaces from compact encodings) and entanglement (enabling capture of higher-order correlations with fewer parameters than purely classical networks).

3. Quantum Kernel Methods: Support Vector Machines and Data Description

Quantum kernel-based models harness feature maps defined by quantum circuits that embed classical input $\mathbf{x}$ into high-dimensional Hilbert spaces. In quantum support vector machines (QSVMs) for anomaly detection, each classical vector $x$ is mapped by a parametric unitary $U_\phi(x)$ , yielding a quantum kernel

$K(x_i, x_j) = |\langle 0^n| U_\phi(x_i)^{\dagger} U_\phi(x_j) | 0^n \rangle|^2.$

Kernels constructed in this way often induce richer nonlinear decision boundaries, as quantum feature maps can lift otherwise inseparable data into linearly separable Hilbert spaces. The Gram matrix is then supplied to a classical SVM solver (e.g., LIBSVM). Quantum kernels can achieve higher kernel-target alignment (e.g., 0.148 vs. 0.077 for classical, a 91% gain) and strongly improved F1-scores (13% higher) over best classical kernels on ICS datasets. Hardware noise studies reveal that moderate quantum kernel errors (up to 0.98%) translate into manageable accuracy loss ( $\sim$ 1.6%) (Cultice et al., 21 Jun 2025). Analogous quantum formulations exist for Support Vector Data Description (SVDD), in which a compact, shallow QCNN ansatz learns the minimum-radius quantum hypersphere enclosing normal data, outperforming quantum autoencoders and classical deep SVDD at constant parameter counts (Oh et al., 2023).

Projected quantum kernels and hardware-efficient feature maps further enable one-class SVM anomaly detection on NISQ processors, with systematic empirical generalization gains and provable margin bounds when qubit-level partial tomography is combined with RBF-like kernels (Pranjić et al., 2024).

4. Quantum Autoencoders and Variational Detection

Quantum autoencoders (QAEs) use parameterized quantum circuits to compress input states into lower-dimensional latent spaces and reconstruct them. Key designs leverage amplitude encoding (mapping a feature vector into quantum amplitudes) and entangling ansätze (e.g., RealAmplitude, PauliTwoDesign), combined with overlap-based loss functions such as direct SWAP tests to measure fidelity between original and reconstructed states. In both time-series (Frehner et al., 2024) and cybersecurity logs (Senthil et al., 22 Oct 2025), QAEs achieve superior anomaly detection metrics—often using two orders of magnitude fewer parameters and fewer iterations than classical autoencoders. Detection is achieved by flagging inputs with high reconstruction error or low overlap with the learned subspace; SWAP test probabilities serve as native quantum anomaly scores.

These advantages derive from the exponential size of Hilbert space reached by few qubits, strong parameter efficiency (e.g., only 21–77 parameters for 7-qubit models), and enhanced expressivity due to quantum entanglement and superposition. Simulations and hardware runs confirm robustness to moderate gate and measurement noise.

5. Quantum Generative Models: GANs and Manifold Learning

Variants of quantum GANs implement both generator and discriminator as variational quantum circuits, or inject quantum-correlated latent priors (e.g., from photonic boson samplers) into otherwise classical GAN pipelines. Successive Data Injection (SuDaI) and data re-uploading are employed to accommodate high-dimensional/time-series data in models with constrained qubit budgets, reusing qubits through circuit depth to implicitly expand feature capacity (Hammami et al., 16 May 2025, Hammami et al., 30 Oct 2025).

Quantum GANs have demonstrated the ability to match or exceed classical GAN anomaly-detection accuracy with an order of magnitude fewer training samples (Bermot et al., 2023). Ensemble GANs with quantum latent spaces outperform classical ensembles in detecting subtle or high-dimensional process anomalies (AUC up to ~0.43 versus 0.35), and real photonic quantum hardware yields nontrivial gains relative to ideal simulation (Kailasanathan et al., 29 Aug 2025). Quantum-enhanced manifold learning with hyperspherical embedding (e.g., Q-BAR) enables robust anomaly recognition in highly data-limited regimes, especially for multi-modal or personalized anomaly tasks (Wang, 11 Dec 2025).

6. Robustness, Scalability, and Advanced Architectures

Partitioned quantum neural networks (PQNNs) address scaling limits in quantum-enhanced anomaly detection. By dividing high-dimensional input into smaller, independently encoded quantum patches, QUPID achieves scalability to smart-grid-scale feature counts and enables practical training and inference within NISQ-era quantum hardware constraints (Ngo et al., 16 Jan 2026). QUPID and its differential privacy extension R-QUPID further demonstrate ML robustness against adversarial attacks, leveraging both quantum superposition and privacy-amplifying quantum noise channels.

Federated quantum learning strategies (PQFL) distribute quantum model training across heterogeneous device fleets, with each client learning a personalized PQC and sharing embeddings following a regularized aggregation protocol. By tuning the personalization regularizer and circuit parameters, PQFL improves anomaly detection AUROC by up to 24 points over classical federated CNNs—especially in non-IID and label-scarce environments (Rahman et al., 8 Nov 2025).

Hybrid unsupervised quantum similarity learning, as applied to LHC new physics searches, deploys Siamese transformer encoders with quantum swap-test similarity heads and explicit noise-mitigating clustering on Bloch-sphere measurement distributions (Hammad et al., 2024). This design achieves AUCs near or exceeding classical baselines and demonstrates empirical resilience to shot-noise and hardware decoherence.

7. Practical Limitations, Hardware Constraints, and Future Directions

Near-term quantum-enhanced anomaly detection is mainly limited by qubit count, circuit fidelity, and the ability to efficiently encode large datasets. Full all-to-all connectivity and dense objective matrices (e.g., in QUBO) are not yet practical on sparse QPU topologies, necessitating constraint-friendly encodings or sparse kernel approximations. Hybrid, partitioned, and federated frameworks offer partial relief, while error mitigation and noise-aware training are essential for successful deployment on NISQ hardware.

Open questions include optimal feature map/ansatz design, scaling quantum kernel and Gram matrix computations, error-resilient readout/tomography protocols, and the systematic integration of quantum circuits within larger, domain-adapted anomaly detection pipelines. The move to real QPU deployment, demonstrated on IBM Falcon/Eagle and photonic processors, marks a substantive advance from pure simulation. As hardware matures, quantum-enhanced anomaly detection is expected to generalize further in domain applicability, sample efficiency, and resilience to complex, high-dimensional, and adversarial data challenges (Mellaerts, 2023, Naaman et al., 19 Sep 2025, Hammami et al., 30 Oct 2025, Cultice et al., 21 Jun 2025, Wang, 11 Dec 2025, Oh et al., 2023, Ngo et al., 16 Jan 2026, Kailasanathan et al., 29 Aug 2025, Bhowmik et al., 2024).