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

Quantum Kernels for Audio Deepfake Detection Using Spectrogram Patch Features

Al Amin, Rakib Hossain, Mahbubul Islam, Faisal Quader, Thanh Thi Nguyen

Published May 7, 2026• 0 citations
Publisher

Abstract

Quantum machine learning has emerged as a promising tool for pattern recognition, yet many audio-focused approaches still treat spectrograms as generic images and do not explicitly exploit their time-frequency structure. We propose Q-Patch, a quantum feature map tailored to audio that encodes local time-frequency patches from mel-spectrograms into quantum states using shallow, hardware-efficient circuits with adjacency-aware entanglement. Each selected patch is summarized by a compact four-dimensional acoustic descriptor and mapped to a four-qubit circuit with depth at most three, enabling practical quantum kernel construction under near-term constraints. We evaluate Q-Patch on an audio spoofing detection task using a controlled, balanced protocol and compare it with size-matched classical baselines. Q-Patch improves discrimination between bona fide and spoofed samples, achieving an area under the receiver operating characteristic curve (AUROC) of 0.87, compared with 0.82 for a radial basis function support vector machine (RBF-SVM) trained on the same patch-level features. Kernel-space analysis further reveals a clear class structure, with cross-class similarity around 0.615 and within-class self-similarity of 1.00. Overall, Q-Patch provides a practical framework for incorporating time-frequency-aware representations into quantum kernel learning for audio authenticity assessment in low-resource settings.

DOI
10.13016/m2tik7-rifu

Related papers

Quantum Kernels for Audio Deepfake Detection Using Spectrogram Patch Features

Lisan Al Amin, Rakib Hossain, Mahbubul Islam +2 more

Quantum machine learning has emerged as a promising tool for pattern recognition, yet many audio-focused approaches still treat spectrograms as generic images and do not explicitly exploit their time-frequency structure. We propose Q-Patch, a quantum feature map tailored to audio that encodes local time-frequency patches from mel-spectrograms into quantum states using shallow, hardware-efficient circuits with adjacency-aware entanglement. Each selected patch is summarized by a compact four-dimensional acoustic descriptor and mapped to a four-qubit circuit with depth at most three, enabling practical quantum kernel construction under near-term constraints. We evaluate Q-Patch on an audio spoofing detection task using a controlled, balanced protocol and compare it with size-matched classical baselines. Q-Patch improves discrimination between bona fide and spoofed samples, achieving an area under the receiver operating characteristic curve (AUROC) of 0.87, compared with 0.82 for a radial basis function support vector machine (RBF-SVM) trained on the same patch-level features. Kernel-space analysis further reveals a clear class structure, with cross-class similarity around 0.615 and within-class self-similarity of 1.00. Overall, Q-Patch provides a practical framework for incorporating time-frequency-aware representations into quantum kernel learning for audio authenticity assessment in low-resource settings.

arXiv:2605.06035•International Joint Conference on Neural Networks (IJCNN)•Jun 1, 2026•1 citation

Quantum Kernel-Based Support Vector Machine for Deepfake Audio Detection

Shivangi Thakur, Mrityunjay Singh, Shatrughan Modi

In this paper, we introduce a quantum kernel-based method to classify deepfake audio using manually extracted acoustic features. Each audio sample is represented by a 194-dimension feature vector. Features are scaled using training-set statistics and compressed to 8 principal components through PCA, then normalized into the range $[-\pi, \pi]$ in order to perform the angle encoding on quantum circuits. Then, we present a quantum feature map with 8-qubits based on IQP-inspired architecture including Hadamard gates, $R Z\left(x_{i}\right)$ rotation and neighboring Ising-ZZ interaction. The quantum kernel is evaluated by the probability of all-zero state when $U \phi\left(x_{i}\right)$ is applied, followed by $U \phi^{\dagger}\left(x_{j}\right)$. We employ the quantum kernel matrix in the SVM classifier to perform real and fake audio detection. For comparison, a classic RBF-SVM and a Hybrid Variational Quantum Classifier trained with the same preprocessed feature representation are evaluated. The experiments were performed on the FoR-2Sec dataset show that QKSVM achieves an accuracy of 0.9894 in accuracy, F1-score of 0.9894, ROC-AUC of 0.9992, EER of 0.0101, and MCC of 0.9788. The McNemar’s test performed on QKSVM and classic RBF-SVM gives $b=41, c=14, \chi^{2}=12.2909, p=0.000455$. This means that the two classifiers have a statistically significant difference at 0.05 level. The results suggest that quantum kernel-based similarity learning can provide a competitive classification approach for deepfake audio detection using compact acoustic representations.

International Conference on Innovative Sustainable Computational Technologies (CISCT)•Jul 1, 2026•0 citations

Empirical Evidence of Quantum Kernel Concentration in Few-Shot Keyword Spotting

Richard Caleb, Jagalingam Pushparaj

Quantum kernel methods utilize high-dimensional quantum feature spaces as a theoretically motivated strategy to address data scarcity in supervised classification. However, theoretical analyses predict exponential concentration of quantum kernel similarity values as qubit count increases. This study presents the first statistically controlled empirical evaluation of fixed-feature-map quantum support vector machines (QSVMs) for few-shot keyword spotting. The ZZFeatureMap and PauliFeatureMap circuits (2 to 3 qubits, PCA-compressed MFCC features, ~51% variance retained) were evaluated across 24 configurations and four low-resource regimes (5–30 samples per class) on the Google Speech Commands dataset, using ten independent seeds with Wilcoxon signed-rank tests and Bonferroni correction. Across all configurations and sample sizes, QSVM accuracy remained close to the five-class random baseline (19.3% at 30 samples per class; random baseline 20%), while RBF SVM achieved 38.7% and a compact CNN achieved 65.3%. Direct kernel similarity measurement confirmed severe concentration, where the 3-qubit QSVM produced near-identical same-class and different-class similarity means (both approximately 0.192), with 94.1% overlap. Under matched MFCC+PCA features, all classical kernels outperformed QSVM, including linear SVM (Cliff’s $\delta = -0.98$ , $p = 0.002$ ) and RBF SVM ( $\delta = -0.82$ , $p = 0.002$ , Bonferroni-corrected $\alpha = 0.003$ ), with CNN achieving complete stochastic dominance ( $\delta = -1.00$ ). These results empirically validate quantum kernel concentration in structured acoustic classification within evaluated NISQ constraints, and apply specifically to the fixed-feature-map paradigm, not quantum kernel methods in general. The identified failure modes, namely exponential concentration, geometric misalignment, and limited circuit expressivity, provide concrete directions for future research into trainable quantum feature maps and hybrid quantum-classical architectures.

IEEE Access•Jan 1, 2026•0 citations

Quantum gate optimization in hybrid neural networks for deepfake audio detection

Abdullah Khan, Wazir Zada Khan, Ayesha Siddiqa +1 more

Audio deepfakes challenge automated authenticity assessment and motivate the investigation of hybrid quantum-classical approaches to speech classification. This study evaluates quantum-circuit design for deepfake audio detection through gate-configuration screening, finite-shot inference analysis, resource estimation, and qubit-scaling experiments. In the exploratory configuration screening of 5 gates, CNOT+S achieved the highest screening score and was subsequently incorporated into a five-qubit Hybrid Quantum-classical Neural Network (HQNN). Across five initialization seeds on DEEP-VOICE, the HQNN achieved \(98.77\%\pm 0.38\%\) accuracy and Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of \(0.9985\pm 0.0011\) , compared with \(95.98\%\pm 0.85\%\) and \(0.9927\pm 0.0014\) , respectively, for a closely parameter-matched classical Convolutional Neural Network (CNN). Zero-shot transfer to ASVspoof 2019 was poor, whereas ASVspoof-specific training produced comparable HQNN and CNN discrimination. Under simulated finite-shot measurement sampling, inference at 100–10,000 shots remained close to analytic expectation-value inference. Logical resource analysis showed rapidly increasing generic amplitude-state-preparation requirements with qubit count, and a controlled five-seed expansion from five to ten qubits did not improve mean classification performance. These results characterize gate-level design, measurement sensitivity, and resource–performance trade-offs in simulated hybrid quantum-classical deepfake audio classification.

Scientific Reports•Nov 12, 2026•0 citations

Quantum Kernel-aided Support Vector Machine Classifier for Improved Speech Classification

Hasmitha Krishna N, S. Sridevi, R Prasanna Kumar

The field of machine learning has been revolutionized by quantum computing, as it has the capacity to tackle complex problems that are beyond the capabilities of classical computers. One such problem is the efficient classification of speech signals with limited data, employing extracted feature attributes like Mel-scale frequency cepstral coefficients (MFCCs). This study compares the performance of classical machine learning models with a Quantum Support Vector Machine (QKSVM) classifier on a data set with a limited number of audio samples. Interestingly, the quantum kernel method outperformed classical models despite the small sample size. This result can be attributed to the quantum kernel method’s ability to extract high-dimensional features efficiently, allowing it to learn more from fewer samples. The Quantum Kernel Circuit (QKC) is contrived in different variants comprising 13, 8, 7, and 4 qubits. This QKC is built with an Angle Embedding layer, an adjoint version of the Angle Embedding layer, and a Hermitian space observable layer. Our findings demonstrate that the hybrid quantum kernel-based technique outperforms existing classical solutions. Amongst variants, the Quantum support vector machine model with 13 optimal numbers of coefficients is efficient in speech classification with the highest accuracy of 87.32%.

International Conference on Computing, Communication and Networking Technologies (ICCCNT)•Jul 6, 2023•14 citations

Benchmarking Quantum Kernel Support Vector Machines Against Classical Baselines on Tabular Data: A Rigorous Empirical Study with Hardware Validation

Siavash Kakavand, Christoph Strohmeyer, Michael Schlotter

Quantum kernel methods have been proposed as a promising approach for leveraging near-term quantum computers for supervised learning, yet rigorous benchmarks against strong classical baselines remain scarce. We present a comprehensive empirical study of quantum kernel support vector machines (QSVMs) across nine binary classification datasets, four quantum feature maps, three classical kernels, and multiple noise models, totalling 970 experiments with strict nested cross-validation. Our analysis spans four phases: (i) statistical significance testing, revealing that none of 29 pairwise quantum-classical comparisons reach significance at $α= 0.05$; (ii) learning curve analysis over six training fractions, showing steeper quantum slopes on six of eight datasets that nonetheless fail to close the gap to the best classical baseline; (iii) hardware validation on IBM ibm_fez (Heron r2), demonstrating kernel fidelity $r \geq 0.976$ across six experiments; and (iv) seed sensitivity analysis confirming reproducibility (mean CV 1.4%). A Kruskal-Wallis factorial analysis reveals that dataset choice dominates performance variance ($\varepsilon^2 = 0.73$), while kernel type accounts for only 9%. Spectral analysis offers a mechanistic explanation: current quantum feature maps produce eigenspectra that are either too flat or too concentrated, missing the intermediate profile of the best classical kernel, the radial basis function (RBF). Quantum kernel training (QKT) via kernel-target alignment yields the single competitive result -- balanced accuracy 0.968 on breast cancer -- but with ~2,000x computational overhead. Our findings provide actionable guidelines for quantum kernel research. The complete benchmark suite is publicly available to facilitate reproduction and extension.

quant-pharXiv:2604.18837•Apr 22, 2026•6 citations
Quantum machine learning has emerged as a promising tool for pattern recognition, yet many audio-focused approaches still treat spectrograms as generic images and do not explicitly exploit their time-frequency structure. We propose Q-Patch, a quantum feature map tailored to audio that encodes local time-frequency patches from mel-spectrograms into quantum states using shallow, hardware-efficient circuits with adjacency-aware entanglement. Each selected patch is summarized by a compact four-dimensional acoustic descriptor and mapped to a four-qubit circuit with depth at most three, enabling practical quantum kernel construction under near-term constraints. We evaluate Q-Patch on an audio spoofing detection task using a controlled, balanced protocol and compare it with size-matched classical baselines. Q-Patch improves discrimination between bona fide and spoofed samples, achieving an area under the receiver operating characteristic curve (AUROC) of 0.87, compared with 0.82 for a radial basis function support vector machine (RBF-SVM) trained on the same patch-level features. Kernel-space analysis further reveals a clear class structure, with cross-class similarity around 0.615 and within-class self-similarity of 1.00. Overall, Q-Patch provides a practical framework for incorporating time-frequency-aware representations into quantum kernel learning for audio authenticity assessment in low-resource settings.
Quantum machine learning has emerged as a promising tool for pattern recognition, yet many audio-focused approaches still treat spectrograms as generic images and do not explicitly exploit their time-frequency structure. We propose Q-Patch, a quantum feature map tailored to audio that encodes local time-frequency patches from mel-spectrograms into quantum states using shallow, hardware-efficient circuits with adjacency-aware entanglement. Each selected patch is summarized by a compact four-dimensional acoustic descriptor and mapped to a four-qubit circuit with depth at most three, enabling practical quantum kernel construction under near-term constraints. We evaluate Q-Patch on an audio spoofing detection task using a controlled, balanced protocol and compare it with size-matched classical baselines. Q-Patch improves discrimination between bona fide and spoofed samples, achieving an area under the receiver operating characteristic curve (AUROC) of 0.87, compared with 0.82 for a radial basis function support vector machine (RBF-SVM) trained on the same patch-level features. Kernel-space analysis further reveals a clear class structure, with cross-class similarity around 0.615 and within-class self-similarity of 1.00. Overall, Q-Patch provides a practical framework for incorporating time-frequency-aware representations into quantum kernel learning for audio authenticity assessment in low-resource settings.
In this paper, we introduce a quantum kernel-based method to classify deepfake audio using manually extracted acoustic features. Each audio sample is represented by a 194-dimension feature vector. Features are scaled using training-set statistics and compressed to 8 principal components through PCA, then normalized into the range [−π,π][-\pi, \pi][−π,π] in order to perform the angle encoding on quantum circuits. Then, we present a quantum feature map with 8-qubits based on IQP-inspired architecture including Hadamard gates, RZ(xi)R Z\left(x_{i}\right)RZ(xi​) rotation and neighboring Ising-ZZ interaction. The quantum kernel is evaluated by the probability of all-zero state when Uϕ(xi)U \phi\left(x_{i}\right)Uϕ(xi​) is applied, followed by Uϕ†(xj)U \phi^{\dagger}\left(x_{j}\right)Uϕ†(xj​). We employ the quantum kernel matrix in the SVM classifier to perform real and fake audio detection. For comparison, a classic RBF-SVM and a Hybrid Variational Quantum Classifier trained with the same preprocessed feature representation are evaluated. The experiments were performed on the FoR-2Sec dataset show that QKSVM achieves an accuracy of 0.9894 in accuracy, F1-score of 0.9894, ROC-AUC of 0.9992, EER of 0.0101, and MCC of 0.9788. The McNemar’s test performed on QKSVM and classic RBF-SVM gives b=41,c=14,χ2=12.2909,p=0.000455b=41, c=14, \chi^{2}=12.2909, p=0.000455b=41,c=14,χ2=12.2909,p=0.000455. This means that the two classifiers have a statistically significant difference at 0.05 level. The results suggest that quantum kernel-based similarity learning can provide a competitive classification approach for deepfake audio detection using compact acoustic representations.
Quantum kernel methods utilize high-dimensional quantum feature spaces as a theoretically motivated strategy to address data scarcity in supervised classification. However, theoretical analyses predict exponential concentration of quantum kernel similarity values as qubit count increases. This study presents the first statistically controlled empirical evaluation of fixed-feature-map quantum support vector machines (QSVMs) for few-shot keyword spotting. The ZZFeatureMap and PauliFeatureMap circuits (2 to 3 qubits, PCA-compressed MFCC features, ~51% variance retained) were evaluated across 24 configurations and four low-resource regimes (5–30 samples per class) on the Google Speech Commands dataset, using ten independent seeds with Wilcoxon signed-rank tests and Bonferroni correction. Across all configurations and sample sizes, QSVM accuracy remained close to the five-class random baseline (19.3% at 30 samples per class; random baseline 20%), while RBF SVM achieved 38.7% and a compact CNN achieved 65.3%. Direct kernel similarity measurement confirmed severe concentration, where the 3-qubit QSVM produced near-identical same-class and different-class similarity means (both approximately 0.192), with 94.1% overlap. Under matched MFCC+PCA features, all classical kernels outperformed QSVM, including linear SVM (Cliff’s δ=−0.98\delta = -0.98δ=−0.98 , p=0.002p = 0.002p=0.002 ) and RBF SVM ( δ=−0.82\delta = -0.82δ=−0.82 , p=0.002p = 0.002p=0.002 , Bonferroni-corrected α=0.003\alpha = 0.003α=0.003 ), with CNN achieving complete stochastic dominance ( δ=−1.00\delta = -1.00δ=−1.00 ). These results empirically validate quantum kernel concentration in structured acoustic classification within evaluated NISQ constraints, and apply specifically to the fixed-feature-map paradigm, not quantum kernel methods in general. The identified failure modes, namely exponential concentration, geometric misalignment, and limited circuit expressivity, provide concrete directions for future research into trainable quantum feature maps and hybrid quantum-classical architectures.
Audio deepfakes challenge automated authenticity assessment and motivate the investigation of hybrid quantum-classical approaches to speech classification. This study evaluates quantum-circuit design for deepfake audio detection through gate-configuration screening, finite-shot inference analysis, resource estimation, and qubit-scaling experiments. In the exploratory configuration screening of 5 gates, CNOT+S achieved the highest screening score and was subsequently incorporated into a five-qubit Hybrid Quantum-classical Neural Network (HQNN). Across five initialization seeds on DEEP-VOICE, the HQNN achieved 98.77%±0.38%98.77\%\pm 0.38\%98.77%±0.38% accuracy and Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.9985±0.00110.9985\pm 0.00110.9985±0.0011 , compared with 95.98%±0.85%95.98\%\pm 0.85\%95.98%±0.85% and 0.9927±0.00140.9927\pm 0.00140.9927±0.0014 , respectively, for a closely parameter-matched classical Convolutional Neural Network (CNN). Zero-shot transfer to ASVspoof 2019 was poor, whereas ASVspoof-specific training produced comparable HQNN and CNN discrimination. Under simulated finite-shot measurement sampling, inference at 100–10,000 shots remained close to analytic expectation-value inference. Logical resource analysis showed rapidly increasing generic amplitude-state-preparation requirements with qubit count, and a controlled five-seed expansion from five to ten qubits did not improve mean classification performance. These results characterize gate-level design, measurement sensitivity, and resource–performance trade-offs in simulated hybrid quantum-classical deepfake audio classification.
The field of machine learning has been revolutionized by quantum computing, as it has the capacity to tackle complex problems that are beyond the capabilities of classical computers. One such problem is the efficient classification of speech signals with limited data, employing extracted feature attributes like Mel-scale frequency cepstral coefficients (MFCCs). This study compares the performance of classical machine learning models with a Quantum Support Vector Machine (QKSVM) classifier on a data set with a limited number of audio samples. Interestingly, the quantum kernel method outperformed classical models despite the small sample size. This result can be attributed to the quantum kernel method’s ability to extract high-dimensional features efficiently, allowing it to learn more from fewer samples. The Quantum Kernel Circuit (QKC) is contrived in different variants comprising 13, 8, 7, and 4 qubits. This QKC is built with an Angle Embedding layer, an adjoint version of the Angle Embedding layer, and a Hermitian space observable layer. Our findings demonstrate that the hybrid quantum kernel-based technique outperforms existing classical solutions. Amongst variants, the Quantum support vector machine model with 13 optimal numbers of coefficients is efficient in speech classification with the highest accuracy of 87.32%.
Quantum kernel methods have been proposed as a promising approach for leveraging near-term quantum computers for supervised learning, yet rigorous benchmarks against strong classical baselines remain scarce. We present a comprehensive empirical study of quantum kernel support vector machines (QSVMs) across nine binary classification datasets, four quantum feature maps, three classical kernels, and multiple noise models, totalling 970 experiments with strict nested cross-validation. Our analysis spans four phases: (i) statistical significance testing, revealing that none of 29 pairwise quantum-classical comparisons reach significance at α=0.05α= 0.05α=0.05; (ii) learning curve analysis over six training fractions, showing steeper quantum slopes on six of eight datasets that nonetheless fail to close the gap to the best classical baseline; (iii) hardware validation on IBM ibm_fez (Heron r2), demonstrating kernel fidelity r≥0.976r \geq 0.976r≥0.976 across six experiments; and (iv) seed sensitivity analysis confirming reproducibility (mean CV 1.4%). A Kruskal-Wallis factorial analysis reveals that dataset choice dominates performance variance (ε2=0.73\varepsilon^2 = 0.73ε2=0.73), while kernel type accounts for only 9%. Spectral analysis offers a mechanistic explanation: current quantum feature maps produce eigenspectra that are either too flat or too concentrated, missing the intermediate profile of the best classical kernel, the radial basis function (RBF). Quantum kernel training (QKT) via kernel-target alignment yields the single competitive result -- balanced accuracy 0.968 on breast cancer -- but with ~2,000x computational overhead. Our findings provide actionable guidelines for quantum kernel research. The complete benchmark suite is publicly available to facilitate reproduction and extension.