Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Journal of Quantum Artificial Intelligence

Out-of-Distribution Detection in Hybrid Quantum Models Using Energy Scores and Measurement Uncertainty

Andi Muhammad Sadat

Published Oct 27, 2026• 0 citations
PDF (CC BY)Publisher

Abstract

Hybrid quantum–classical models deployed in open-world settings require reliable out-of-distribution detection under finite-shot measurement noise, where identical inputs can induce materially different logits across repeated circuit executions. This study evaluates three scoring rules, namely softmax confidence, energy scoring, and uncertainty-aware energy scoring, using a consistent validation-threshold protocol and multi-source OOD testing. Across three OOD shift families, uncertainty-aware energy achieved source-wise AUROC values of 0.953 (semantic shift), 0.942 (corruption shift), and 0.936 (background shift), compared with 0.941, 0.923, and 0.916 for energy-only scoring, and 0.902, 0.884, and 0.864 for softmax confidence. Aggregated over sources, uncertainty-aware energy reached AUROC 0.944 and AUPR 0.931, reducing FPR95 to 0.152, while energy-only produced AUROC 0.927, AUPR 0.914, and FPR95 0.193, and softmax yielded AUROC 0.883, AUPR 0.861, and FPR95 0.279. In-distribution classification remained stable with accuracy 0.898–0.902 across scoring variants, indicating that detection gains were not driven by degraded classification performance. Measurement-budget analysis shows that uncertainty-aware scoring substantially improves low-shot reliability. At 64 shots per observable, energy-only achieved AUROC 0.894 with FPR95 0.251, whereas uncertainty-aware energy reached AUROC 0.929 with FPR95 0.171, while the energy-score dispersion decreased monotonically from 0.42 at 64 shots to 0.17 at 1024 shots. As shot budgets increased, energy-only performance improved, but uncertainty-aware energy retained a consistent advantage, indicating that variance penalization improves score ordering, not only score variance. Architecture ablations further identify a practical expressivity–stability tradeoff: an intermediate-scale circuit (8 qubits, depth 6, brickwork entanglement) attained AUROC 0.949 with FPR95 0.149 and in-distribution accuracy 0.903, whereas further scaling (10 qubits, depth 6) produced AUROC 0.946 with higher FPR95 0.154 and increased score dispersion. Calibration behavior improved concurrently, with ECE reduced from 0.061 (softmax) to 0.044 (energy) and 0.036 (uncertainty-aware energy), supporting more stable operating points under threshold transfer. These findings demonstrate that integrating measurement uncertainty into energy-based scoring provides a robust control mechanism for hybrid quantum OOD detection, particularly under constrained shot budgets and measurement noise.

DOI
10.63913/jqai.v1i3.18

Related papers

Statistical Mechanics-Inspired Quantum Generative Models for Rare-Event Simulation and Anomaly Detection

Rajesh Bhaskarla

Reliable estimation of tail risk and detection of low-prevalence anomalies remain challenging under distribution shift and limited simulation budgets. This paper introduces a statistical mechanics-inspired hybrid quantum generative model that learns an energy-shaped proposal for rare-event simulation while producing a calibrated anomaly score for deployment. Across simulation budgets from 200 to 6400 samples, relative error decreased from 0.31 to 0.08, outperforming normalizing flows (0.44 to 0.12), diffusion sampling (0.48 to 0.14), and a classical energy model (0.46 to 0.13). At a fixed evaluation budget, the proposed estimator achieved relative error 0.10 with effective sample size 1120 and tail hit rate 6.8%, compared with 0.16, 760, and 4.1% for flows. For anomaly detection, AUROC reached 0.962 and AUPRC reached 0.714, with precision 0.61 at 80% recall, exceeding flow scoring (AUROC 0.935, AUPRC 0.624, precision 0.49) and VAE reconstruction (AUROC 0.928, AUPRC 0.598, precision 0.46). Under temporal drift after a single calibration, realized false positive rate stayed within 0.0004 of a 0.5% target, whereas baselines drifted by up to 0.0084. Ablations showed peak AUPRC near moderate circuit depth and mid-scale qubits, while monotone annealing reduced rare-event error from 0.12 to 0.10 and improved stability. Under increasing noise, mitigation preserved AUPRC from 0.74 at low noise to 0.69 at moderate noise, supporting feasibility for NISQ-constrained monitoring and tail estimation.

Journal of Quantum Artificial Intelligence•Aug 31, 2026•0 citations

Uncertainty-Calibrated Quantum Neural Networks via Bayesian Variational Circuits and Conformal Prediction

Bryan Daniel Angelino

Reliable decision-making with quantum neural networks on NISQ hardware requires uncertainty estimates that remain meaningful under finite-shot measurement noise and distribution shift. This study introduces an end-to-end pipeline that combines Bayesian variational quantum circuits with post hoc probability calibration and conformal prediction to produce coverage-controlled prediction sets. Across shot budgets from 128 to 4096 shots, the Bayesian variational circuit improves accuracy from 0.828 to 0.839 at 128 shots and from 0.862 to 0.892 at 4096 shots, while reducing negative log-likelihood from 0.462 to 0.448 at 128 shots and from 0.392 to 0.357 at 4096 shots. Calibration preserves accuracy while further lowering negative log-likelihood to 0.432 at 128 shots and 0.350 at 4096 shots, and reduces expected calibration error from 0.058 to 0.027. Conformalization achieves empirical coverage of 0.952 at miscoverage 0.05, with average set size 1.39 and singleton rate 0.71, and maintains coverage between 0.944 and 0.948 under feature corruption and subpopulation shift, where deterministic accuracy degrades to 0.801 and 0.774, respectively, while the Bayesian model retains 0.842 and 0.816. Ablations identify an operational knee at circuit depth 8, and show diminishing returns beyond 8 posterior samples per prediction, supporting deployment-efficient uncertainty control for quantum learning.

Journal of Quantum Artificial Intelligence•Aug 31, 2026•0 citations

Quantum Autoencoders for Network Anomaly Detection: A Comparative Benchmark with Classical Methods

Gönül ALTAY, Ecir Uğur Küçüksille

The practical value of quantum machine learning for network anomaly detection remains unclear under controlled and comparable benchmarking conditions. This study evaluates a trash-qubit-based Quantum Autoencoder (QAE) against seven classical anomaly-detection methods on UNSW-NB15 using shared data partitions, preprocessing steps, model selection procedures, and threshold calibration protocols. The experimental design included five independent seeds, three training set sizes, two feature budgets, and 12 QAE architectures, together with low-FPR analysis, finite-shot and synthetic-noise evaluation, less-constrained classical reference settings, and a cross-dataset assessment on TON_IoT. The Angle-RY encoding with a Real-Amplitudes Ring ansatz at depth L = 2 achieved the highest mean validation-selection AUPRC. On UNSW-NB15, the QAE was competitive in AUPRC and AUROC but remained behind the strongest classical baselines in F1-score, and its attack recall declined sharply under low-FPR constraints. The TON_IoT results further showed that relative model performance depended on the dataset. Under finite-shot and synthetic-noise conditions, anomaly-score rankings were largely preserved, whereas threshold-based decisions were more sensitive. Overall, the QAE showed competitive continuous-score discrimination on UNSW-NB15, but achieving reliable detection at low false-alarm rates remains a key operational challenge.

Electronics•Oct 1, 2026•0 citations

Q-CaMIR: Uncertainty-Aware Hybrid Quantum–Classical Learning for robust Cross-Modal radiological image classification

Ali Jaber Almalki, Heba Kadry, Hanin Ardah +4 more

Achieving reliable radiological image classification requires not only high accuracy, but also well-calibrated uncertainty estimates, robustness in the face of domain shifts, and generalizability across various imaging modalities. Current deep learning methodologies for medical image classification frequently display overconfident predictions and experience significant degradation under realistic clinical conditions, such as label noise, limited training data, and distribution shifts between institutions. Hybrid quantum–classical machine learning has emerged as a promising paradigm; however, existing studies in medical imaging predominantly focus on single-modality and accuracy-centric approaches. In this study, we introduced Q-CaMIR (Cross-Modal, Uncertainty-Aware Hybrid Quantum–Classical Learning), a novel framework that integrates chest X-ray and brain MRI classification within a unified hybrid quantum–classical architecture, incorporating explicit uncertainty quantification. Q-CaMIR utilizes a pretrained classical vision backbone to extract modality-agnostic latent representations, aligns cross-modal features through a contrastive representation module, and channels the resulting compact embeddings through a variational quantum circuit (VQC) decision head, enhanced with evidential deep learning for uncertainty-aware prediction. We evaluated Q-CaMIR on three public benchmarks: CheXpert and MIMIC-CXR (external validation) for chest radiograph classification, and BraTS 2021 for brain tumor classification, employing a comprehensive suite of metrics including accuracy, macro-F1, AUROC, AUPRC, expected calibration error (ECE), and Brier score. Extensive experiments under low-data regimes, synthetic label noise, intensity corruption, and cross-institutional domain shifts demonstrate that Q-CaMIR achieves competitive classification performance compared to robust classical baselines while significantly improving calibration (up to 47% ECE reduction) and robustness. Ablation studies confirm the complementary contributions of the quantum decision head and uncertainty-aware training objective. Our findings suggest that hybrid quantum–classical learning provides a meaningful inductive bias for uncertainty-sensitive radiological classification, advancing towards clinically deployable AI systems for radiation-related diagnostic workflows.

Journal of Radiation Research and Applied Sciences•Jun 1, 2026•0 citations

Hybrid Quantum–Classical Convolutional Networks for Robust Multimodal Pattern Recognition

Qingmei Li

Robust multimodal pattern recognition remains challenging under real-world conditions characterized by sensor degradation, distribution shift, and missing modalities. This study introduces a Hybrid Quantum–Classical Convolutional Network (Hybrid QC–CNN) that combines modality-specific convolutional encoders with a compact Parameterized Quantum Circuit (PQC) feature transformation and mask-aware fusion to stabilize inference under partial and corrupted evidence. Under matched-capacity baselines, the proposed model achieves 0.913 clean accuracy and 0.907 macro-F1 while delivering superior robustness across corruption severities. Averaged over severity levels 1–5, Hybrid QC–CNN attains 0.812 accuracy compared to 0.782 (multimodal transformer) and 0.771 (CNN late fusion), and improves worst-severity (level 5) accuracy to 0.701, exceeding the transformer (0.602) by +9.9 percentage points and late fusion (0.583) by +11.8 points. Robustness summarized by AURC reaches 0.823, surpassing the transformer (0.793) and late fusion (0.781). Reliability under degradation is also improved: ECE is 0.032 on clean data and 0.061 at severity 3, compared to 0.089 (transformer) and 0.094 (late fusion). Missing-modality stress tests further show stable degradation: with only one modality available, Hybrid QC–CNN achieves 0.793 (image-only), 0.742 (audio-only), and 0.764 (text-only), outperforming both transformer and late-fusion baselines across all single-modality conditions. Ablation removing the quantum layer reduces average corruption accuracy from 0.812 to 0.792 and worst-severity accuracy from 0.701 to 0.641, indicating that the PQC contributes meaningful robustness beyond mask-aware fusion and regularization. Collectively, results support hybrid quantum–classical feature transformation as a targeted mechanism for multimodal robustness and calibrated confidence.

Journal of Quantum Artificial Intelligence•May 1, 2026•0 citations

Adaptive-Shot Hybrid Quantum Anomaly Detection for Tactile Internet Security: Reliability-Aware Measurement Allocation Under Resource Constraints

Mubassir Serneabat Sudipto, Shakil Ahmed, Ashfaq Khokhar +1 more

Tactile Internet (TI) security analytics must balance reliable thresholded decisions with constrained computational and measurement resources. We study this tension for finite-shot hybrid quantum anomaly inference and introduce the Adaptive-Shot Variational Quantum Circuit (AS-VQC) policy. This validation-calibrated policy begins each record at 128 shots and cumulatively escalates through 256, 512, and 1024 shots only when the finite-shot anomaly score remains close to a validation-selected security threshold. The quantum scorer is evaluated as an off-path security analytics component rather than part of the haptic critical path. Using a 4,875-record CESNET-TimeSeries24-derived aggregate-flow benchmark, leakage-safe random, entity-group-disjoint, and temporal holdouts, and five trained quantum neural network (QNN) checkpoints per holdout, the primary AS-VQC-95 (beta = 0.95) policy averages 129.2, 276.9, and 131.2 shots per record, saving 87.4%, 73.0%, and 87.2% of the uniform 1024-shot baseline (Fixed-1024), respectively. The decision disagreement with analytic (exact-expectation) inference is 0.771%, 0.409%, and 0.635%, lower than both the uniform 128-shot baseline (Fixed-128) and a matched-budget shuffled-allocation control. Fixed-1024 remains more decision-stable, establishing a measurable reliability-resource trade-off rather than cost-free equivalence. A more conservative AS-VQC-99 (beta = 0.99) further reduces disagreement while using fewer than 512 average shots across all holdouts. These results show that finite quantum measurements can be treated as an inference resource and concentrated on boundary-sensitive TI-security decisions while exposing checkpoint-dependent escalation under unseen-entity conditions.

arXiv:2610.05835•Oct 5, 2026•0 citations
Hybrid quantum–classical models deployed in open-world settings require reliable out-of-distribution detection under finite-shot measurement noise, where identical inputs can induce materially different logits across repeated circuit executions. This study evaluates three scoring rules, namely softmax confidence, energy scoring, and uncertainty-aware energy scoring, using a consistent validation-threshold protocol and multi-source OOD testing. Across three OOD shift families, uncertainty-aware energy achieved source-wise AUROC values of 0.953 (semantic shift), 0.942 (corruption shift), and 0.936 (background shift), compared with 0.941, 0.923, and 0.916 for energy-only scoring, and 0.902, 0.884, and 0.864 for softmax confidence. Aggregated over sources, uncertainty-aware energy reached AUROC 0.944 and AUPR 0.931, reducing FPR95 to 0.152, while energy-only produced AUROC 0.927, AUPR 0.914, and FPR95 0.193, and softmax yielded AUROC 0.883, AUPR 0.861, and FPR95 0.279. In-distribution classification remained stable with accuracy 0.898–0.902 across scoring variants, indicating that detection gains were not driven by degraded classification performance. Measurement-budget analysis shows that uncertainty-aware scoring substantially improves low-shot reliability. At 64 shots per observable, energy-only achieved AUROC 0.894 with FPR95 0.251, whereas uncertainty-aware energy reached AUROC 0.929 with FPR95 0.171, while the energy-score dispersion decreased monotonically from 0.42 at 64 shots to 0.17 at 1024 shots. As shot budgets increased, energy-only performance improved, but uncertainty-aware energy retained a consistent advantage, indicating that variance penalization improves score ordering, not only score variance. Architecture ablations further identify a practical expressivity–stability tradeoff: an intermediate-scale circuit (8 qubits, depth 6, brickwork entanglement) attained AUROC 0.949 with FPR95 0.149 and in-distribution accuracy 0.903, whereas further scaling (10 qubits, depth 6) produced AUROC 0.946 with higher FPR95 0.154 and increased score dispersion. Calibration behavior improved concurrently, with ECE reduced from 0.061 (softmax) to 0.044 (energy) and 0.036 (uncertainty-aware energy), supporting more stable operating points under threshold transfer. These findings demonstrate that integrating measurement uncertainty into energy-based scoring provides a robust control mechanism for hybrid quantum OOD detection, particularly under constrained shot budgets and measurement noise.
Reliable estimation of tail risk and detection of low-prevalence anomalies remain challenging under distribution shift and limited simulation budgets. This paper introduces a statistical mechanics-inspired hybrid quantum generative model that learns an energy-shaped proposal for rare-event simulation while producing a calibrated anomaly score for deployment. Across simulation budgets from 200 to 6400 samples, relative error decreased from 0.31 to 0.08, outperforming normalizing flows (0.44 to 0.12), diffusion sampling (0.48 to 0.14), and a classical energy model (0.46 to 0.13). At a fixed evaluation budget, the proposed estimator achieved relative error 0.10 with effective sample size 1120 and tail hit rate 6.8%, compared with 0.16, 760, and 4.1% for flows. For anomaly detection, AUROC reached 0.962 and AUPRC reached 0.714, with precision 0.61 at 80% recall, exceeding flow scoring (AUROC 0.935, AUPRC 0.624, precision 0.49) and VAE reconstruction (AUROC 0.928, AUPRC 0.598, precision 0.46). Under temporal drift after a single calibration, realized false positive rate stayed within 0.0004 of a 0.5% target, whereas baselines drifted by up to 0.0084. Ablations showed peak AUPRC near moderate circuit depth and mid-scale qubits, while monotone annealing reduced rare-event error from 0.12 to 0.10 and improved stability. Under increasing noise, mitigation preserved AUPRC from 0.74 at low noise to 0.69 at moderate noise, supporting feasibility for NISQ-constrained monitoring and tail estimation.
Reliable decision-making with quantum neural networks on NISQ hardware requires uncertainty estimates that remain meaningful under finite-shot measurement noise and distribution shift. This study introduces an end-to-end pipeline that combines Bayesian variational quantum circuits with post hoc probability calibration and conformal prediction to produce coverage-controlled prediction sets. Across shot budgets from 128 to 4096 shots, the Bayesian variational circuit improves accuracy from 0.828 to 0.839 at 128 shots and from 0.862 to 0.892 at 4096 shots, while reducing negative log-likelihood from 0.462 to 0.448 at 128 shots and from 0.392 to 0.357 at 4096 shots. Calibration preserves accuracy while further lowering negative log-likelihood to 0.432 at 128 shots and 0.350 at 4096 shots, and reduces expected calibration error from 0.058 to 0.027. Conformalization achieves empirical coverage of 0.952 at miscoverage 0.05, with average set size 1.39 and singleton rate 0.71, and maintains coverage between 0.944 and 0.948 under feature corruption and subpopulation shift, where deterministic accuracy degrades to 0.801 and 0.774, respectively, while the Bayesian model retains 0.842 and 0.816. Ablations identify an operational knee at circuit depth 8, and show diminishing returns beyond 8 posterior samples per prediction, supporting deployment-efficient uncertainty control for quantum learning.
The practical value of quantum machine learning for network anomaly detection remains unclear under controlled and comparable benchmarking conditions. This study evaluates a trash-qubit-based Quantum Autoencoder (QAE) against seven classical anomaly-detection methods on UNSW-NB15 using shared data partitions, preprocessing steps, model selection procedures, and threshold calibration protocols. The experimental design included five independent seeds, three training set sizes, two feature budgets, and 12 QAE architectures, together with low-FPR analysis, finite-shot and synthetic-noise evaluation, less-constrained classical reference settings, and a cross-dataset assessment on TON_IoT. The Angle-RY encoding with a Real-Amplitudes Ring ansatz at depth L = 2 achieved the highest mean validation-selection AUPRC. On UNSW-NB15, the QAE was competitive in AUPRC and AUROC but remained behind the strongest classical baselines in F1-score, and its attack recall declined sharply under low-FPR constraints. The TON_IoT results further showed that relative model performance depended on the dataset. Under finite-shot and synthetic-noise conditions, anomaly-score rankings were largely preserved, whereas threshold-based decisions were more sensitive. Overall, the QAE showed competitive continuous-score discrimination on UNSW-NB15, but achieving reliable detection at low false-alarm rates remains a key operational challenge.
Achieving reliable radiological image classification requires not only high accuracy, but also well-calibrated uncertainty estimates, robustness in the face of domain shifts, and generalizability across various imaging modalities. Current deep learning methodologies for medical image classification frequently display overconfident predictions and experience significant degradation under realistic clinical conditions, such as label noise, limited training data, and distribution shifts between institutions. Hybrid quantum–classical machine learning has emerged as a promising paradigm; however, existing studies in medical imaging predominantly focus on single-modality and accuracy-centric approaches. In this study, we introduced Q-CaMIR (Cross-Modal, Uncertainty-Aware Hybrid Quantum–Classical Learning), a novel framework that integrates chest X-ray and brain MRI classification within a unified hybrid quantum–classical architecture, incorporating explicit uncertainty quantification. Q-CaMIR utilizes a pretrained classical vision backbone to extract modality-agnostic latent representations, aligns cross-modal features through a contrastive representation module, and channels the resulting compact embeddings through a variational quantum circuit (VQC) decision head, enhanced with evidential deep learning for uncertainty-aware prediction. We evaluated Q-CaMIR on three public benchmarks: CheXpert and MIMIC-CXR (external validation) for chest radiograph classification, and BraTS 2021 for brain tumor classification, employing a comprehensive suite of metrics including accuracy, macro-F1, AUROC, AUPRC, expected calibration error (ECE), and Brier score. Extensive experiments under low-data regimes, synthetic label noise, intensity corruption, and cross-institutional domain shifts demonstrate that Q-CaMIR achieves competitive classification performance compared to robust classical baselines while significantly improving calibration (up to 47% ECE reduction) and robustness. Ablation studies confirm the complementary contributions of the quantum decision head and uncertainty-aware training objective. Our findings suggest that hybrid quantum–classical learning provides a meaningful inductive bias for uncertainty-sensitive radiological classification, advancing towards clinically deployable AI systems for radiation-related diagnostic workflows.
Robust multimodal pattern recognition remains challenging under real-world conditions characterized by sensor degradation, distribution shift, and missing modalities. This study introduces a Hybrid Quantum–Classical Convolutional Network (Hybrid QC–CNN) that combines modality-specific convolutional encoders with a compact Parameterized Quantum Circuit (PQC) feature transformation and mask-aware fusion to stabilize inference under partial and corrupted evidence. Under matched-capacity baselines, the proposed model achieves 0.913 clean accuracy and 0.907 macro-F1 while delivering superior robustness across corruption severities. Averaged over severity levels 1–5, Hybrid QC–CNN attains 0.812 accuracy compared to 0.782 (multimodal transformer) and 0.771 (CNN late fusion), and improves worst-severity (level 5) accuracy to 0.701, exceeding the transformer (0.602) by +9.9 percentage points and late fusion (0.583) by +11.8 points. Robustness summarized by AURC reaches 0.823, surpassing the transformer (0.793) and late fusion (0.781). Reliability under degradation is also improved: ECE is 0.032 on clean data and 0.061 at severity 3, compared to 0.089 (transformer) and 0.094 (late fusion). Missing-modality stress tests further show stable degradation: with only one modality available, Hybrid QC–CNN achieves 0.793 (image-only), 0.742 (audio-only), and 0.764 (text-only), outperforming both transformer and late-fusion baselines across all single-modality conditions. Ablation removing the quantum layer reduces average corruption accuracy from 0.812 to 0.792 and worst-severity accuracy from 0.701 to 0.641, indicating that the PQC contributes meaningful robustness beyond mask-aware fusion and regularization. Collectively, results support hybrid quantum–classical feature transformation as a targeted mechanism for multimodal robustness and calibrated confidence.
Tactile Internet (TI) security analytics must balance reliable thresholded decisions with constrained computational and measurement resources. We study this tension for finite-shot hybrid quantum anomaly inference and introduce the Adaptive-Shot Variational Quantum Circuit (AS-VQC) policy. This validation-calibrated policy begins each record at 128 shots and cumulatively escalates through 256, 512, and 1024 shots only when the finite-shot anomaly score remains close to a validation-selected security threshold. The quantum scorer is evaluated as an off-path security analytics component rather than part of the haptic critical path. Using a 4,875-record CESNET-TimeSeries24-derived aggregate-flow benchmark, leakage-safe random, entity-group-disjoint, and temporal holdouts, and five trained quantum neural network (QNN) checkpoints per holdout, the primary AS-VQC-95 (beta = 0.95) policy averages 129.2, 276.9, and 131.2 shots per record, saving 87.4%, 73.0%, and 87.2% of the uniform 1024-shot baseline (Fixed-1024), respectively. The decision disagreement with analytic (exact-expectation) inference is 0.771%, 0.409%, and 0.635%, lower than both the uniform 128-shot baseline (Fixed-128) and a matched-budget shuffled-allocation control. Fixed-1024 remains more decision-stable, establishing a measurable reliability-resource trade-off rather than cost-free equivalence. A more conservative AS-VQC-99 (beta = 0.99) further reduces disagreement while using fewer than 512 average shots across all holdouts. These results show that finite quantum measurements can be treated as an inference resource and concentrated on boundary-sensitive TI-security decisions while exposing checkpoint-dependent escalation under unseen-entity conditions.