The DOI matches its Crossref record; 45 of the 45 references it cites were found in a registry or an index; 10 of the 11 sources its claims cite were checked for support.
Quantum-Trained Convolutional Neural Network for Deepfake Audio Detection
Open accessAbstract
The rise of deepfake technologies has posed significant challenges to privacy, security, and information integrity, particularly in audio and multimedia content. This paper introduces a Quantum-Trained Convolutional Neural Network (QT-CNN) framework designed to enhance the detection of deepfake audio, leveraging the computational power of quantum machine learning (QML). The QT-CNN employs a hybrid quantum-classical approach, integrating Quantum Neural Networks (QNNs) with classical neural architectures to optimize training efficiency while reducing the number of trainable parameters. Our method incorporates a novel quantum-to-classical parameter mapping that effectively utilizes quantum states to enhance the expressive power of the model, achieving up to 70% parameter reduction compared to classical models without compromising accuracy. Data pre-processing involved extracting essential audio features, label encoding, feature scaling, and constructing sequential datasets for robust model evaluation. Experimental results demonstrate that the QT-CNN achieves comparable performance to traditional CNNs, maintaining high accuracy during training and testing phases across varying configurations of QNN blocks.
Overview
CoChat Summary
TL;DR
The paper proposes a Quantum-Trained Convolutional Neural Network (QT-CNN) for deepfake audio detection, using a hybrid quantum-classical scheme to generate CNN weights from quantum neural network measurement probabilities. On the DEEP-VOICE dataset, it reports accuracy comparable to a classical CNN while reducing trainable parameters, with claims of up to 70% parameter reduction.
Key findings
- The QT-CNN maintained high accuracy during training and testing across varying QNN block configurations, closely aligning with classical CNN performance despite fewer trainable parameters.
- The classical CNN had 3,373 parameters; the QT model’s total trainable parameters ranged from 456 to 1,464 as QNN blocks increased from 12 to 96, corresponding to 13.5%–43.4% of the classical model’s parameters.
- The abstract and conclusion claim up to 70% parameter reduction without compromising accuracy, and the results text states parameter efficiency reaches up to 70% at 96 QNN blocks.
- The QNN blocks used 12 qubits; the mapping model had 301 parameters and the scaling model had 8 parameters.
Methods
- Hybrid quantum-classical QT-CNN: an N-qubit QNN with parameterized Ry gates and linear CNOT entanglement; a quantum-to-classical mapping model maps measurement probabilities to CNN weights; training uses cross-entropy loss with gradients computed analytically or via the parameter-shift rule.
- Data: the DEEP-VOICE dataset, described as balanced real and AI-generated deepfake audio; extracted features included chroma STFT, spectral centroid, spectral bandwidth, roll-off frequency, zero-crossing rate, and MFCCs; preprocessing used LabelEncoder, MinMaxScaler, sliding windows, and stratified train/validation/test splits.
- Evaluation compared the QT-CNN against a classical CNN with two convolutional and two fully connected layers across varying QNN block counts from 12 to 96.
Limitations
- The paper does not state explicit limitations.
- Evident from the text: no exact accuracy values or statistical tests are reported; performance is described qualitatively as comparable or high.
- Evident from the text: parameter-reduction claims are inconsistent—the abstract/conclusion say up to 70% reduction, while the reported ratio at 96 blocks is 43.4% of classical parameters (56.6% reduction)—and evaluation is limited to the DEEP-VOICE dataset.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 0 Not valid · 0 Not checked · 19 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“scaling down from M parameters in classical NNs to O ( polylog ( M )) by using QNNs [21].”
- Also stated in: § C. Training Procedure for Quantum-Enhanced Learning · p. 2 · plain
“reducing QNN parameters to O ( polylog ( M )) , compared to M for classical NNs”
What the sources say
- Supportedcites [21] Liu, 2024, Quantum-train: Rethinking hybrid quantum-classical machine learning in the model compression perspective source
The Introduction states that QT reduces trainable parameters to O(polylog(M)), where M is the number required by a classical NN. Section C, “Training Flow for Quantum-Train,” explicitly contrasts O(polylog(M)) QNN parameters with M classical-NN parameters. (LLM, conf=0.95)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § C. Training Procedure for Quantum-Enhanced Learning · p. 2 · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Our approach employs a hybrid quantum-classical training mechanism, where quantum circuits are utilized to optimize the weight parameters of the CNN, thus enhancing its performance and robustness against adversarial inputs [19].”
What the sources say
- Partly supportedcites [19] sourceExperimental — accuracy not yet measured
The ARCHITECTURE section (page unavailable) discusses hybrid models that pass data through quantum filters before a classical CNN and says they may inherit adversarial-robustness guarantees. It presents that robustness as a possibility and an open question, not as a demonstrated improvement in a QT-CNN; it also does not say quantum circuits optimize CNN weights. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This methodology not only addresses the inherent limitations of classical CNNs in detecting deepfake content but also circumvents some of the scaling issues faced by purely quantum models in the Noisy IntermediateScale Quantum (NISQ) era [20].”
What the sources say
- Partly supportedcites [20] sourceExperimental — accuracy not yet measured
The Introduction (p. 3) describes the limited, noisy qubits available in the NISQ era, and “Scope of the review” (p. 6) explains how hybrid quantum–classical algorithms divide computation between quantum and classical devices. These passages support the background motivation, but they do not establish that the citing paper’s QT framework circumvents scaling issues faced by purely quantum models. (LLM, conf=0.83)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“DEEPVOICE [26] comprises both authentic human speech and AIgenerated deepfake audio, designed to simulate realistic scenarios of voice manipulation.”
What the sources say
- Supportedcites [26] Bird, 2023, Real-time detection of ai-generated speech for deepfake voice conversion source
The Abstract states that DEEP-VOICE contains real human speech and speech converted between speakers using Retrieval-based Voice Conversion. Section 3.1 describes preserving ambient sounds while converting the speaker’s voice, consistent with the citing sentence’s description of realistic voice manipulation. (LLM, conf=0.91)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The total number of trainable parameters in the QT model varies with the number of QNN blocks, ranging from 456 to 1,464 parameters as the number of QNN blocks increases from 12 to 96.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
12 more claims
Create a free account to see all 26 claims and every source verdict.
- Probably not a substantial claim. 6
- It may state two separate claims as one. 5
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Quantum-Trained Convolutional Neural Network for Deepfake Audio Detection · p. 1
“achieving up to 70% parameter reduction compared to classical models without compromising accuracy.”
§ C. Classification Result and Analysis · p. 4
“This corresponds to a parameter ratio ranging from 13.5% to 43.4% relative to the classical model”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C. Classification Result and Analysis · p. 4
“revealing a clear trend of increased parameter efficiency as the number of QNN blocks grows, reaching up to 70% efficiency at 96 blocks.”
§ C. Classification Result and Analysis · p. 4
“This corresponds to a parameter ratio ranging from 13.5% to 43.4% relative to the classical model”
Citations
Cites
Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., & Lloyd, S. (2017) · Nature · doi:10.1038/nature23474 · Cited by 4,779 · cited 1×
“QML, an emerging field that combines the computational power of quantum mechanics with the versatility of machine learning, has shown promising potential in various domains [6]–[17].”
Noisy intermediate-scale quantum algorithms
Bharti, K., Cervera-Lierta, A., Kyaw, T. H., Haug, T., Alperin-Lea, S., Anand, A., Degroote, M., Heimonen, H., Kottmann, J. S., Menke, T., et al. (2022) · Reviews of Modern Physics · doi:10.1103/revmodphys.94.015004 · Cited by 2,231
Quantum Machine Learning in Feature Hilbert Spaces
Schuld, M., & Killoran, N. (2019) · Physical Review Letters · doi:10.1103/physrevlett.122.040504 · Cited by 1,840
The Emergence of Deepfake Technology: A Review
Westerlund, M. (2019) · Technology Innovation Management Review · doi:10.22215/timreview/1282 · Cited by 1,138 · cited 1×
“The rapid advancement of machine learning (ML) has paved the way for innovative applications, but it has also introduced challenges, particularly in the realm of synthetic media, such as deepfake audio and video [1].”
The power of quantum neural networks
Abbas, A., Sutter, D., Zoufal, C., Lucchi, A., Figalli, A., & Woerner, S. (2021) · Nature Computational Science · doi:10.1038/s43588-021-00084-1 · Cited by 1,082 · cited 1×
“Quantum neural networks (QNNs), for instance, exploit quantum superposition and entanglement to process information in ways that classical networks cannot, potentially accelerating training times and improving performance on complex data sets [18].”
Power of data in quantum machine learning
Huang, H.-Y., Broughton, M., Mohseni, M., Babbush, R., Boixo, S., Neven, H., & McClean, J. R. (2021) · Nature Communications · doi:10.1038/s41467-021-22539-9 · Cited by 896 · cited 1×
“By leveraging quantum-enhanced techniques, QML can offer unique advantages in scenarios where large-scale data processing and intricate pattern recognition are essential [7].”
Data re-uploading for a universal quantum classifier
Pérez-Salinas, A., Cervera-Lierta, A., Gil-Fuster, E., & Latorre, J. I. (2020) · Quantum · doi:10.22331/q-2020-02-06-226 · Cited by 713
Variational Quantum Circuits for Deep Reinforcement Learning
Chen, S. Y.-C., Yang, C.-H. H., Qi, J., Chen, P.-Y., Ma, X., & Goan, H.-S. (2020) · IEEE Access · doi:10.1109/access.2020.3010470 · Cited by 452
Self-citationDeepfake Audio Detection via MFCC features using Machine Learning
Hamza, A., Javed, A. R. R., Iqbal, F., Kryvinska, N., Almadhor, A. S., Jalil, Z., & Borghol, R. (2022) · IEEE Access · doi:10.1109/access.2022.3231480 · Cited by 264 · cited 2×
“…consisted of extracted audio features such as chroma STFT [27], spectral centroid [28], spectral bandwidth [29], roll-off frequency, zero-crossing rate [30], and Mel-Frequency Cepstral Coefficients (MFCCs) [3], which were read into a pandas DataFrame for exploratory data analysis (EDA) [31].”
Quantum Long Short-Term Memory
Chen, S. Y.-C., Yoo, S., & Fang, Y.-L. L. (2022) · ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp43922.2022.9747369 · Cited by 233
Self-citation
Research works printed but not linked: 1
- Müller, 2015, Short-time fourier transform and chroma features
Cited by
Differentiable Quantum Architecture Search in Quantum-Enhanced Neural Network Parameter Generation
Chen, S. Y.-C., Liu, C.-Y., Chen, K., Huang, W., Chang, Y.-J., & Huang, W. (2025) · IEEE International Conference on Quantum Computing and Engineering (QCE) · doi:10.1109/qce65121.2025.00170 · Cited by 15
“In image classification, QT-based architectures have been applied to build compact yet expressive visual models [39], [42], [43].”
Self-citationQuantum-Enhanced Parameter-Efficient Learning for Typhoon Trajectory Forecasting
Liu, C.-Y., Chen, K., Chen, Y., Chen, S. Y.-C., Huang, W., Huang, W., & Chang, Y.-J. (2025) · IEEE International Conference on Quantum Computing and Engineering (QCE) · doi:10.1109/qce65121.2025.00224 · Cited by 13
“Current studies on QT have demonstrated promising results across various domains, including image classification [35], [37], reinforcement learning [38]–[40], long short-term memory (LSTM) models and regression tasks [41]–[43].”
Self-citationIntroduction to Quantum Machine Learning and Quantum Architecture Search
Chen, S. Y.-C., & Liang, Z. (2025) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas56072.2025.11044136 · Cited by 8
“Empirically, hybrid approaches have been successfully applied to various domains, including classification [4]–[7], time-series prediction [8], natural language processing [9]– [12], reinforcement learning [13]–[21], and neural network model compression [22]–[26].”
Self-citationA Survey on Quantum Machine Learning in Speech Acoustics
Kucharski, M. A. (2025) · IEEE International Symposium on Computational Intelligence and Informatics (CINTI) · doi:10.1109/cinti67731.2025.11311733 · Cited by 5
Distributed Quantum Neural Networks on Distributed Photonic Quantum Computing
Chen, K., Liu, C.-Y., Yu, S., Burt, F., & Leung, K. K. (2025) · IEEE International Conference on Quantum Computing and Engineering (QCE) · doi:10.1109/qce65121.2025.00165 · Cited by 7
“Quantum parameter updates occur asynchronously across photonic nodes, with classical HPC aggregating gradients via consensus protocols, [36].”
Self-citation
Metadata
- Authors
- Lin, C.-H. A., Liu, C.-Y., Chen, S. Y.-C., & Chen, K.-C.
- Affiliations
- Imperial College London (GB)
- National Taiwan University (TW)
- Wells Fargo (United States) (US)
- Venue
- 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW)
- Publisher
- IEEE
- Published
- 2025-04-06
- Pages
- 1-5
- OpenAlex
- W4410771607
- Cited by
- 30
- References
- 35
- Citations by year
- 2026: 13 · 2025: 17
- FWCI
- 28.45
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW) (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Funders
- Engineering and Physical Sciences Research Council
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Quantum Computing Algorithms and Architecture
- Keywords
- audio deepfake detection, quantum machine learning, quantum neural networks, hybrid quantum-classical neural networks, convolutional neural networks, parameter reduction, audio feature extraction
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Quantum-Trained Convolutional Neural Network for Deepfake Audio Detection. https://app.cochat.ai/papers/10.1109/icasspw65056.2025.11011133
- The paper
- Lin, C.-H. A., Liu, C.-Y., Chen, S. Y.-C., & Chen, K.-C. (2025). Quantum-Trained Convolutional Neural Network for Deepfake Audio Detection. 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). https://doi.org/10.1109/icasspw65056.2025.11011133
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.