Hybrid Classical Quantum Learning Model Framework for Detection of
Deepfake Audio
Atul Pandey
a
and Bhawana Rudra
b
Department of Information Technology, National Institute of Technology,
Srinivasnagar-575025, Mangalore, Dakshina Kannada, Karnataka, India
Keywords:
Quantum Machine Learning, Quantum Deep Learning, Quantum-Deepfake Audio, Hybrid Classical Quantum
Model, Deepfake Audio.
Abstract:
Artificial intelligence (AI) has simplified individual tasks compared to earlier times. However, it also enables
the creation of fake images, audio, and videos that can be misused to tarnish the reputation of a person on
social media. The rapid advancement of deepfake technology presents significant challenges in detecting
such fabricated content. Therefore, in this paper, we particularly focus on the deepfake audio detection.
Many Classical models exist to detect deepfake audio, but they often overlook critical audio features, and
training these models can be computationally resource-intensive. To address this issue, we used a real-time
AI-generated fake speech dataset, which includes all the necessary features required to train models and used
Quantum Machine Learning (QML) techniques, which follow principles of quantum mechanics to process
the data simultaneously. We propose a hybrid Classical-Quantum Learning Model that takes advantage of
Classical and Quantum Machine Learning. The hybrid model is trained on a real-time AI-generated fake
speech dataset, and we compare the performance with existing Classical and Quantum models in this area.
Our results show that the hybrid Classical-Quantum model gives an accuracy of 98.81% than the Quantum
Support vector Machine (QSVM) and Quantum Neural Network (QNN).
1 INTRODUCTION
Technological innovation continues to simplify hu-
man tasks, with one key advancement being Artificial
Intelligence (AI), which enables people to work more
efficiently and intelligently. AI can be used to gen-
erate images, videos, digital avatars, and even video
dubbing (Nguyen et al., 2022). Unfortunately, this
technology is sometimes exploited to tarnish the rep-
utation of individuals by creating fake content, such
as forged voices, images, and videos, using deep-
fake techniques—where deep learning models are
employed to fabricate such content (Dagar and Vish-
wakarma, 2022). In a recent case, fraudsters uti-
lized AI-driven software to imitate the voice of a
company’s CEO, successfully extorting USD 243,000
(The Wall Street Journal, 2019). As a result, there is
growing interest in developing methods for detecting
fraudulent voices (Khochare et al., 2021). People of-
ten rely on their knowledge and environmental aware-
ness to identify fake audio. However, the rapid ad-
a
https://orcid.org/0009-0004-3106-4995
b
https://orcid.org/0000-0001-7651-3820
vancements in deepfake audio generation have under-
scored the importance of addressing these challenges.
One specific type of deepfake audio is voice conver-
sion, where the voice of a person is swapped with an-
other (Yi et al., 2023).
Researchers commonly rely on Classical Deep-
Learning models to identify deepfake content. How-
ever, training these models demands significant
computational resources, even when using high-
performance hardware like Graphics Processing Units
(GPUs). While GPUs offer parallel processing capa-
bilities, their performance is limited by the number
of cores, leading to slower processing speeds than
Quantum computers. Quantum computers leverage
the principles of Quantum mechanics—such as super-
position, entanglement, and interference, to process
data much more efficiently and faster. These systems
operate on Quantum bits (Qubits), each representing
a superposition of Quantum states. Qubits provide
exponential computational speed by simultaneously
accessing multiple Quantum states. The processing
power of the system is determined by the number of
Qubits (denoted as N), with the ability to access 2
N
states concurrently (Zaman et al., 2024).
Pandey, A. and Rudra, B.
Hybrid Classical Quantum Learning Model Framework for Detection of Deepfake Audio.
DOI: 10.5220/0013258700003899
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 11th International Conference on Information Systems Security and Privacy (ICISSP 2025) - Volume 2, pages 231-239
ISBN: 978-989-758-735-1; ISSN: 2184-4356
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
231

Figure 1: Four Sub-Categories of QML (A
¨
ımeur et al.,
2006).
Quantum Machine Learning (QML) offers a range
of methods to address complex problems efficiently.
Problems in QML can be approached using four dis-
tinct methods, as illustrated in figure 1. The four sub-
categories of the QML use the Quantum-inspired ma-
chine learning algorithm to process the Classical or
Quantum data involved in the problem, and data pro-
cessing is achieved on either a Classical or Quantum
computer (A
¨
ımeur et al., 2006). Each sub-categories
are outlined below :
• CC Approach. Classical data is processed on a
Classical computer using a Quantum-inspired ma-
chine learning algorithm.
• CQ Approach. Classical data is processed on
a Quantum computer using a Quantum-inspired
machine learning algorithm.
• QC Approach. Quantum data is processed on a
Classical computer using a Quantum-inspired ma-
chine learning algorithm.
• QQ Approach. Quantum data is processed on
a Quantum computer using a Quantum-inspired
machine learning algorithm.
This paper is structured as follows: Section 2 dis-
cusses the existing Classical and Quantum models in
the deepfake audio. This section also discusses the
background of Quantum circuits. Section 3 provides
a method to detect the deepfake audio and also dis-
cusses the proposed hybrid model. Section 4 provides
the comparative analysis of both systems. Section 5
provides the insights of the implementation and the
overview of the model performance. Section 6 pro-
vides a detailed analysis of the paper. Section 7 con-
cludes the paper and discusses future work.
2 LITERATURE REVIEW
Several research papers have explored deepfake audio
classification using machine learning and deep learn-
ing models. The author proposed a Convolutional
Neural Network (CNN) based classifier, such as Light
CNN, that filters noise in voice signals while preserv-
ing key information (Wu et al., 2018). Convolutional-
Recurrent Neural Network (CRNN) based spoofing
detection uses five 1D convolutional layers, a Long
Short-Term Memory (LSTM) Layer, and two fully
connected Layers to perform end-to-end detection of
deepfake audio (Chintha et al., 2020). The authors
(Khochare et al., 2021) proposed two approaches:
one uses audio features for classification via ma-
chine learning models, while the other classifies us-
ing a temporal convolutional network and spatial
transformer network based on images of the audio
signals. While deep learning models achieve bet-
ter results, they did not consider Short-Time Fourier
Transform (STFT) and Mel-Frequency Cepstral Co-
efficients (MFCCs), which are two of the most ef-
fective features of the audio signal. The authors
(Zhang et al., 2021) proposed a Squeeze and Exci-
tation Network (SENet) that captures interdependen-
cies between channels but requires more computa-
tional time for training. The authors (Hamza et al.,
2022) presented a method for handling large datasets
and classifying them using various machine learning
algorithms. The Support Vector Machine (SVM) per-
formed well on the For-rece and For-2-sec datasets,
but for the For-norm dataset, gradient boosting gen-
erates better results. However, this work did not ad-
dress fluctuations and distortions in the audio signals.
The authors (Mcuba et al., 2023) extracted the vari-
ous features from the fake audio file, such as MFCCs,
Mel-Spectrum, Chromagram, and Spectrogram and
converted them into images. The custom model and
VGG model were trained on the audio features, and
the results show that VGG performed well for the
MFCCs feature and the custom model performed bet-
ter for the rest of the features. The authors (Doan
et al., 2023) proposed a breathing-talking-silence en-
coder to detect deepfake audio using ASVspoof 2019
and 2021 datasets. The results show that the perfor-
mance of the classifier increased by 40%. The authors
(Wu et al., 2024) proposed a deepfake detector based
on contrastive Learning. This method minimized the
variation in audio, which happened because of the
manipulation of audio. This will increase the robust-
ness of the model for the detection of deepfake audio.
The author (Pham et al., 2024) used an ASVspoof
2019 benchmark dataset and extracted the Spectro-
gram from the audio. The CNN-based model, var-
ious pre-trained models and ensemble models were
trained on the Spectrogram. The results show that the
ensemble model performs better than the other mod-
els. The authors (Li et al., 2024) proposed a SafeEar
framework to detect deepfake audio without relying
on semantic content such that private content remains
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
232

secure in audio. They introduced the neural audio
codec that separates the semantic and acoustic infor-
mation, and they rely only on the acoustic informa-
tion. The framework was tested on the 4 datasets and
shows an error rate down to 2.02%, which made it
suitable for anti-deepfake and anti-content recovery.
However, the proposed method is limited to acous-
tic features, which makes it less effective against nu-
anced manipulations mimicking natural patterns. The
authors (Saha et al., 2024) proposed a method to ex-
ecute the machine learning and deep learning pro-
gram for deepfake audio detection on the Central Pro-
cessing Unit (CPU). This framework utilizes the self-
supervised learning-based pre-trained model. The re-
sults show that the author achieved the 0.90% error
rate with 1000 trainable parameters. Several papers
focus on Classical models, and only a few have ex-
plored Quantum learning models for the deepfake im-
age. Therefore, this paper focuses on the deepfake
audio.
The authors (Mittal et al., 2020) proposed a
method to detect fake images based on feature ex-
traction using a Quantum-inspired evolutionary algo-
rithm, though it lacks fine-tuning of parameters and
noise filtering. The authors (Mishra and Samanta,
2022) introduced a Quantum-based transfer learning
approach to detect deepfake images, where features
are extracted from a pre-trained ResNet-18 model and
classified using a Quantum Neural Network (QNN).
The authors (Pandey and Rudra, 2024) proposed a
method to detect deepfake audio speech using a Quan-
tum Support Vector Machine (QSVM) and QNN.
However, the performance of the model was not better
compared to Classical.
The challenge to detect deepfake audio lies in rec-
ognizing features within the audio signal, whether
they are genuine or fake. Literature shows that AI-
based models are capable of effectively learning and
predicting audio authenticity. These models are based
on Classical Deep-Learning techniques. Many au-
thors proposed the Quantum model to take Quantum
advantage to overcome issues in a Classical computer.
However, the Quantum model for detecting deepfake
audio did not perform well compared to the Clas-
sical model. Therefore, to improve the model per-
formance, we propose a hybrid Classical-Quantum
learning model that takes advantage of Classical and
Quantum Machine Learning.
2.1 Quantum Preliminaries
A Quantum circuit comprises two essential compo-
nents: the feature map and the variational form. The
feature map encodes Classical data into a Quantum
Figure 2: Parametrized Quantum Circuit (Benedetti et al.,
2019).
state, while the variational form adjusts this Quantum
state to the desired target state by iteratively tuning
parameters.
• Feature Maps. There are several methods for
embedding Classical data into Quantum states
through feature mapping. Common techniques in-
clude Angle Embedding.
– Angle Embedding. Angle Embedding is
one of the simplest approaches for encoding
floating-point data. It transforms a single
floating-point value x ∈ R into a Quantum state
using the following equation (1):
R
k
(x)|0⟩ = e
−ix
σ
k
2
|0⟩ (1)
Here, k ∈ {x, y, z} represents the rotation axis
on the Bloch sphere, implemented through
Pauli rotation gates. These rotations are applied
to the data being encoded. In the case of An-
gle Embedding, the number of rotations corre-
sponds to the number of features in the dataset
(Schuld and Petruccione, 2018).
• Parametrized Quantum Circuits (PQCs). Vari-
ational Quantum Circuits (VQCs) or Parametrized
Quantum Circuits (PQCs) are quantum algorithms
by their reliance on free parameters. In QML,
VQCs encode the Classical data into a Quantum
state using the feature maps discussed in section
2.1 and then perform a variational form to create
the QNN. The parameters used in the variational
form are optimized through an iterative process.
Measurement is performed on a Quantum circuit,
which leads to stochastic output. We repeat the
experiment multiple times to get the expectation
value, and this will result in a probability distribu-
tion of the basis states. This probability distribu-
tion is given to the Classical algorithm to compute
the loss function or cost function, which gives the
difference between the predicted and true labels.
These results are given to the Classical optimizer
to update the parameters of the Quantum circuit
to minimize the loss function. Figure 2 shows the
working principle of the PQCs (Benedetti et al.,
2019), which consists of feature mapping, varia-
tional forms and optimisation.
Hybrid Classical Quantum Learning Model Framework for Detection of Deepfake Audio
233

Figure 3: Proposed Methodology.
Algorithm 1: TwoLocal(n,k, θ).
Input: n, k, θ
for r = 0 to k do
▷ Add the r-th layer.;
for j = 1 to n do
Apply a R
Y
(θ
r j
) gate on qubit j.;
end
▷ Create entanglement between layers.;
if r < k then
for t = 1 to n − 1 do
Apply a CNOT gate with control on
qubit t and target on qubit t + 1.;
end
end
end
– Variational Form. The variational form of a
Quantum Neural Network (QNN) mimics the
layered architecture of Classical Neural Net-
works. It relies on optimizable parameters
−→
θ
and introduces entanglement between Qubits
through a parameter-independent circuit U
t
ent
.
Multiple Layers (or repetitions) can be stacked
in the variational circuit. In our model, we em-
ploy a TwoLocal variational form which uses n
Qubits and k repetitions, the total number of pa-
rameters needed for optimization is n×(k + 1).
These parameters, denoted as θ
r j
, are indexed
by r (from 0 to k) and j (from 1 to n). It will
create k Layers not the k +1 Layers. Algorithm
1 defines the creation of the TwoLocal varia-
tional form (El
´
ıas Fern
´
andez, 2023).
• Quantum Kernel. Consider a Quantum model
f (x) defined as in equation (2):
f (x) = ⟨ψ(x)|M|ψ(x)⟩ (2)
In equation 2, |ψ(x)⟩ is a Quantum state gener-
ated by an embedding circuit that encodes the in-
put data x, and M is a chosen observable. ⟨ψ(x)| is
the transpose of the |ψ(x)⟩ i.e ⟨ψ(x)| = (|ψ(x)⟩)
†
.
This formulation encompasses variational QML
models because the observable M can be realized
through a simple measurement, which is preceded
by a variational circuit. Instead of training the
function f using variational methods, we can of-
ten achieve the same result by employing a Clas-
sical kernel method, where the kernel is computed
on a Quantum device. The equation (3) shows
the Quantum kernel determined by the overlap be-
tween two Quantum states encoding different data
points:
κ(x,x
′
) = |⟨ψ(x
′
)|ψ(x)⟩|
2
(3)
By using this kernel-based approach, we avoid
the need for processing and measuring the typi-
cal variational circuits, focusing solely on the data
encoding (Schuld and Killoran, 2018).
3 METHODOLOGY
(Pandey and Rudra, 2024) proposed deepfake speech
detection using Quantum models such as QSVM
and QNN to compare the performance with Classi-
cal models such as Support Vector Machine (SVM)
and Artificial Neural Networks (ANN). To improve
the detection performance of the Quantum model,
we proposed a hybrid Classical-Quantum model that
takes advantage of Classical and Quantum Machine
Learning. We also train the classical 1D CNN to com-
pare it with the proposed model. A numerical dataset
is utilized to train the models. The dataset contains
features from the audio speech. Figure 3 illustrates
the proposed methodology, which is broken down into
three stages:
• Input Phase. In this stage, the dataset is taken as
input and undergoes preprocessing.
• Training and Testing Phase. This stage takes the
Classical data and encodes it into Quantum states
using the embedding technique, and the varia-
tional form is applied to create the QNN (refer to
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
234

Figure 4: Hybrid Classical 1D Convolution Quantum Neural Network.
Figure 5: Quantum Layer Circuit of Hybrid Model
(HC1CQNN).
section 2.1). The hybrid model is trained, and its
performance is evaluated using the test dataset.
• Output Phase. This stage compares the results
from both the Classical models, Quantum models,
and the hybrid model.
3.1 Dataset
We consider the recent audio numerical dataset ”Real-
time detection of AI-Generated speech for Deepfake”,
published in 2023. J. J. Bird and A. Lotfi applied
the two-stem model (Hennequin et al., 2020) from
Spleeter to separate actual speech into natural vo-
cals and accompaniment (background noise). The
Spleeter model comprises 12 Layers organized into
two sets of 6 Layers each for the encoder-decoder
Convolutional Neural Network (CNN) within a U-
Net architecture. Following this, the unprocessed vo-
cals were converted into synthesized vocals of dis-
tinct individuals utilizing the Retrieval-Based Voice
Conversion (RVC) model. Subsequently, the back-
ground noise and RVC-generated vocals were amal-
gamated to produce synthetic speech. The author has
employed the Python-based Librosa library (McFee
et al., 2015) to extract the 26 different features. The
features include the chromagram (chromastft), spec-
tral centroid, spectral bandwidth, spectral rolloff, root
mean square (RMS) and twenty Mel-Frequency Cep-
stral Coefficients (MFCCs) (Bird and Lotfi, 2023).
3.2 Preprocessing
To ensure that the model does not become biased
towards any particular class, the dataset should be
shuffled to introduce randomness. It is divided into
80% for training and 20% for testing, which helps
to prevent data leakage. Since some features in the
dataset have varying ranges, feature scaling is applied
to bring them to a uniform scale. This scaling im-
proves the algorithm’s convergence speed. The Min-
Max scaling method normalises the features while
preserving the original data range and enhancing in-
terpretability. The scaler function is fitted on the train-
ing data and then applied to transform the test data,
thereby avoiding data leakage during model evalua-
tion.
3.3 Hybrid Classical Quantum
Learning Model
Literature shows that CNN helps the model to learn
necessary features from the data, which is extracted
using the convolution operation. To take advantage
Hybrid Classical Quantum Learning Model Framework for Detection of Deepfake Audio
235

of the CNN, we propose a Hybrid Classical 1D Con-
volution Quantum Neural Network (HC1CQNN) as
shown in figure 4. This hybrid model combines 4
Layers- 1D Convolution Layer, Classical Neurons
Layer, Quantum Layer, and Single Classical Neuron
Layer. The Classical 1D Convolution extracts the im-
portant feature from the audio dataset and feeds the
features as input to the Classical Neurons. This will
reduce the dimension of the data, which will be help-
ful for the Quantum Layer because we can not feed
all the extracted features into the Quantum Layer be-
cause of the limitations of the current Quantum Sim-
ulators. The end Layer (Single Classical Neuron) of
the hybrid model will classify the audio speech as real
or fake. The hybrid model is trained as a single unit.
The hybrid model in figure 4 performs the con-
volution operation on the data of size 1x26 with 32
filters of size 3, which results in the 32 features map
of size 1x26, and then applies the Max Pooling Layer
(size-2) on the feature maps, which produces the 32
feature maps of size 1x13. Each feature map contains
13 features, which generates 416 features after flat-
tening. 416 features cannot directly feed as input to
the Quantum Layer because of the restriction of the
qubit and the limitation of Quantum simulators. To
handle this issue, we have applied 32 Classical Neu-
rons followed by 4 Classical Neurons after flatten-
ing the features, and their output will be input to the
Quantum Layer with 4 qubits to avoid system crash.
The Quantum Layer converts the Classical data into
a Quantum state using Angle embedding, which then
learns the pattern in the data after embedding using
the TwoLocal variational forms algorithm discussed
in section 2.1 and then performs measurements on all
the qubits. Figure 5 shows the Quantum Layer circuit
diagram of HC1CQNN, which includes Angle em-
bedding, TwoLocal variational forms and measure-
ment of the circuit. The measurement result will be
the input for the Single Classical Neuron, which later
performs the classification of fake and real audio.
4 QUANTUM AND CLASSICAL
SYSTEM ANALYSIS
This section analyzes the Quantum system Q and
Classical system C to evaluate their computational
performance in deep learning tasks. Let us assume
that both systems take classical data x consisting of n
bits as input.
x ← Classical data
n ← Number of bits
The system Q processes x using the PQC dis-
cussed in section 2.1. This system processes all 2
n
states simultaneously. Let us assume that Q requires
p units to process these 2
n
states, from encoding to
measurement. The measurement results are then fed
into a classical optimizer, which adjusts the parame-
ters used in Q. Assume the optimizer takes q units per
optimization iteration. Thus, the PQC requires (p+q)
units for a single run. To reach the desired minimum
loss, the PQC runs z times.
p ← Units to process Q
q ← Units taken by the classical optimizer
p + q ← Units taken by PQC for a single run
z ← Times to run PQC
The total processing time for system Q is:
T
Q
= (p + q) × z (4)
Now, let’s assume the Classical system C with
GPU can handle m states concurrently. With 2
n
states
to process, C would need to perform approximately
(2
n
/m) sequential processing steps. Each processing
step requires r units, and its result is fed into the clas-
sical optimizer, which adjusts the parameters for C.
Thus, system C requires
2
n
m
×(r + q) units for a single
run. To achieve the minimum loss, system C also runs
z times.
m ← States handle by the GPU
2
n
m
← Sequential processing steps
r ← Units required per processing step
2
n
m
× (r + q) ← Units taken by system C for a single run
z ← Times to run system C
The total processing time for the Classical system
C with GPU is:
T
GPU
C
=
2
n
m
× (r + q) × z (5)
For the Quantum system Q to outperform the
Classical system C with GPU support, we require:
(p + q) × z <
2
n
m
× (r + q) × z
Dividing by z (assuming z ̸= 0) gives:
p + q <
2
n
m
× (r + q)
Based on this analysis, we observe that the Quan-
tum system Q maintains an advantage as n grows
larger. As 2
n
grows exponentially,
2
n
m
× (r + q) be-
comes large even with significant GPU parallelism.
Thus, while adding GPU with Classical system C, it
still does not eliminate the exponential scaling chal-
lenge faced by Classical processing.
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
236

Figure 6: Loss Graph of 1D Convolution Neural Network
During Training.
Figure 7: Loss Graph of Hybrid Classical 1D Convolution
Quantum Neural Network During Training.
5 IMPLEMENTATION &
RESULTS
The authors (Bird and Lotfi, 2023) created a .csv file
comprising both real and fake voice samples. This
file contains 11,778 data points (rows) and 26 features
(columns). The authors (Pandey and Rudra, 2024)
subsequently reduced the dataset to 2,000 data points
and applied Principal Component Analysis (PCA) to
reduce the feature count from 26 to 13 for training
the Quantum models (QSVM and QNN) to prevent
system crashes. We have utilized the entire dataset
to enhance the performance of the Quantum model
through a hybrid model approach. We employed an
NVIDIA RTX A4000 GPU to run the classical and hy-
brid models. Both the hybrid model and the Classical
1D CNN were trained on all 11,778 data points with
26 features to ensure a fair comparison of the models
on the same scale. The Classical 1D CNN and the hy-
brid model were implemented using TensorFlow ver-
sion 2.15.0. For the hybrid model, we used Penny-
lane (Bergholm et al., 2022) version 0.36.0 to encode
the classical data and construct the Quantum Layer
using Pennylane’s TensorFlow interface. The hy-
brid model training was conducted on the lightning
qubit simulator provided by Pennylane, which of-
fers efficient linear algebra computation and differ-
entiation methods to train the hybrid model effec-
tively. These simulators use Quantum algorithms to
leverage Quantum properties to execute QML pro-
grams. However, at the hardware level, simulators
run on Classical computers, which may require sev-
eral days or even weeks to complete QML tasks, re-
sulting in more execution time and sometimes sys-
tem crashes. Table 1 presents the classification met-
rics for both Classical (SVM and ANN) and Quantum
(QSVM and QNN) models, as discussed (Pandey and
Rudra, 2024). 90.02%, 95.97%, 83.50% and 89.30%
are the accuracy, precision, recall and f1-score respec-
tively represent the performance metric of the QSVM.
70.07%, 72.47%, 64.50% and 68.25% are the accu-
racy, precision, recall, and f1-score respectively rep-
resent the performance metric of the QNN. These
results indicate that the reduced dataset and simula-
tor limitation leads to performance degradation for
QSVM and QNN. Our hybrid model implementation
overcomes these issues and yields improved results.
Table 2 provides the classification metrics for the
models—1D CNN and hybrid model (HC1CQNN) on
the test dataset. We obtain the training loss graph as
shown in figure 6 and 7. We observe that all the pa-
rameters range from 98-99% for both 1D CNN and
hybrid model (HC1CQNN). we observe that the hy-
brid model is trained perfectly and is almost similar
to the training loss graph of the Classical 1D CNN
model. This shows that the hybrid model has im-
proved its performance over the Quantum model.
6 DISCUSSION
In this study, we evaluate both Classical and Quantum
Machine Learning for deepfake audio detection. To
leverage the Quantum advantage, the author (Pandey
and Rudra, 2024) uses the Quantum models to per-
form deepfake audio detection. However, the per-
formance of the Quantum model lags behind that of
the classical model, as in table 1. This likely hap-
pened because of the use of fewer features and data
(rows) from dataset (Bird and Lotfi, 2023) as well as
the limitation of simulators. Therefore to improve the
detection performance, we have proposed the hybrid
model. This model leverages the advantage of the
Classical and Quantum models. The results in table
2 show that the hybrid model achieved almost simi-
lar results compared to Classical 1D CNN. This in-
dicates that the performance of the Quantum model
Hybrid Classical Quantum Learning Model Framework for Detection of Deepfake Audio
237

Table 1: Classification Metric Results of Classical and Quantum Model.
Model Type Models Accuracy
(%)
Precision
(%)
Recall (%) F1-score (%)
Classical
Model
Classical SVM
(Pandey and Rudra,
2024)
83.79 86.43 81.90 84.10
Classical ANN
(Pandey and Rudra,
2024)
95.00 96.27 93.29 94.76
Quantum
Model
QSVM (Pandey and
Rudra, 2024)
90.02 95.97 83.50 89.30
QNN (Pandey and
Rudra, 2024)
70.07 72.47 64.50 68.25
Table 2: Classification Metric Results of Classical and Hybrid Model.
Model Type Models Accuracy
(%)
Precision
(%)
Recall (%) F1-score (%)
Classical Model Classical 1D CNN 99.19 98.89 99.48 99.19
Quantum Model Hybrid Model
(HC1CQNN)
98.81 98.39 99.23 98.80
(QSVM and QNN) can be enhanced through a hy-
brid approach. Our analysis in section 4 suggests that
Quantum systems offer faster computation than Clas-
sical systems. However, in practice, this advantage is
not purely realized due to the current limitations of
Quantum simulators.
7 CONCLUSION AND FUTURE
WORK
This paper demonstrates the application of Quan-
tum models for deepfake audio detection, utilizing
the computational advantages offered by Quantum
processing. Quantum approaches were considered,
as Classical computers encounter significant compu-
tational challenges, particularly with the extensive
resources required for training deep learning mod-
els. However, literature shows that Quantum models
are not up to when compared with Classical models
due to the use of fewer features and data from the
dataset as well as the limitation of Quantum simula-
tors. Therefore, to improve the performance of Quan-
tum models (QSVM and QNN), we propose a hybrid
approach that leverages the strengths of both Classi-
cal and Quantum models. The results indicate that the
hybrid model performs almost similar to the Classical
1D CNN model. Our analysis shows that Quantum
systems have the potential to perform faster computa-
tions than Classical systems. However, this advantage
remains constrained in practice due to the limitations
of current Quantum simulators. Deploying deepfake
audio detection using the Quantum model effectively
requires large datasets and improved Quantum sim-
ulators to prevent systems from crashing. With the
existing technology, our hybrid model demonstrates
improved performance with a combination of Clas-
sical and Quantum Machine Learning techniques.
However, achieving optimal performance with purely
Quantum models will require further development in
Quantum simulators. In the future, we will explore a
hybrid Classical-Quantum Model approach for other
areas of deepfake detection, such as video deepfake
detection.
REFERENCES
A
¨
ımeur, E., Brassard, G., and Gambs, S. (2006). Machine
learning in a quantum world. In Lamontagne, L. and
Marchand, M., editors, Advances in Artificial Intelli-
gence, pages 431–442, Berlin, Heidelberg. Springer
Berlin Heidelberg.
Benedetti, M., Lloyd, E., Sack, S., and Fiorentini, M.
(2019). Parameterized quantum circuits as machine
learning models. Quantum Science and Technology,
4(4):043001.
Bergholm, V., Izaac, J., Schuld, M., Gogolin, C., Ahmed,
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
238

S., and et.al., V. A. (2022). Pennylane: Automatic
differentiation of hybrid quantum-classical computa-
tions.
Bird, J. J. and Lotfi, A. (2023). Real-time detection of ai-
generated speech for deepfake voice conversion.
Chintha, A., Thai, B., and et.al., S. (2020). Recurrent con-
volutional structures for audio spoof and video deep-
fake detection. IEEE Journal of Selected Topics in
Signal Processing, 14:1024–1037.
Dagar, D. and Vishwakarma, D. (2022). A literature review
and perspectives in deepfakes: generation, detection,
and applications. International Journal of Multimedia
Information Retrieval, 11.
Doan, T.-P., Nguyen-Vu, L., Jung, S., and Hong, K. (2023).
Bts-e: Audio deepfake detection using breathing-
talking-silence encoder. ICASSP 2023 - 2023 IEEE
International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pages 1–5.
El
´
ıas Fern
´
andez, Combarro
´
Alvarez, S. G. C. (2023). A
Practical Guide to Quantum Machine Learning and
Quantum Optimization. Packt, 1st edition.
Hamza, A., Javed, A. R. R., Iqbal, F., Kryvinska, N., Al-
madhor, A. S., Jalil, Z., and Borghol, R. t. (2022).
Deepfake audio detection via mfcc features using ma-
chine learning. IEEE Access, 10:134018–134028.
Hennequin, R., Khlif, A., Voituret, F., and Moussallam, M.
(2020). Spleeter: a fast and efficient music source
separation tool with pre-trained models. Journal of
Open Source Software, 5(50):2154.
Khochare, J., Joshi, C., Yenarkar, B., Suratkar, S., and Kazi,
F. t. (2021). A deep learning framework for audio
deepfake detection. Arabian Journal for Science and
Engineering, 47.
Li, X., Li, K., Zheng, Y., Yan, C., Ji, X., and et.al, W. X.
(2024). Safeear: Content privacy-preserving audio
deepfake detection.
McFee, B., Raffel, C., and et.al., L. (2015). librosa: Audio
and music signal analysis in python. In Proceedings
of the 14th python in science conference, volume 8,
pages 18–25.
Mcuba, M., Singh, A., Ikuesan, R. A., and et.al., H. V.
(2023). The effect of deep learning methods on deep-
fake audio detection for digital investigation. Pro-
cedia Computer Science, 219:211–219. CENTERIS
– International Conference on ENTERprise Informa-
tion Systems / ProjMAN – International Conference
on Project MANagement / HCist – International Con-
ference on Health and Social Care Information Sys-
tems and Technologies 2022.
Mishra, B. and Samanta, A. (2022). Quantum transfer
learning approach for deepfake detection. Sparkling-
light Transactions on Artificial Intelligence and Quan-
tum Computing.
Mittal, H., Saraswat, M., Bansal, J. C., and Nagar, A.
(2020). Fake-face image classification using im-
proved quantum-inspired evolutionary-based feature
selection method. In 2020 IEEE Symposium Series on
Computational Intelligence (SSCI), pages 989–995.
Nguyen, T. T., Nguyen, Q. V. H., Nguyen, D. T., Nguyen,
D. T., Huynh-The, T., Nahavandi, S., Nguyen, T. T.,
Pham, Q.-V., and Nguyen, C. M. t. (2022). Deep
learning for deepfakes creation and detection: A sur-
vey. Computer Vision and Image Understanding,
223:103525.
Pandey, A. and Rudra, B. (2024). Deepfake audio detection
using quantum learning models. In Proceedings of
the IEEE Middle East Conference on Communications
and Networking.
Pham, L., Lam, P., Nguyen, T., Nguyen, H., and
Schindler, A. (2024). Deepfake audio detection us-
ing spectrogram-based feature and ensemble of deep
learning models.
Saha, S., Sahidullah, M., and Das, S. (2024). Exploring
green ai for audio deepfake detection.
Schuld, M. and Killoran, N. (2018). Quantum machine
learning in feature hilbert spaces. Physical review let-
ters, 122 4:040504.
Schuld, M. and Petruccione, F. (2018). Supervised Learn-
ing with Quantum Computers. Springer Publishing
Company, Incorporated, 1st edition.
The Wall Street Journal (2019). Fraudsters use ai to mimic
ceo’s voice in unusual cybercrime case. Accessed on
May 28, 2024.
Wu, H., Chen, J., Du, R., Wu, C., He, K., Shang, X., Ren,
H., and Xu, G. (2024). Clad: Robust audio deep-
fake detection against manipulation attacks with con-
trastive learning.
Wu, X., He, R., and et.al., S. (2018). A light cnn for
deep face representation with noisy labels. IEEE
Transactions on Information Forensics and Security,
13(11):2884–2896.
Yi, J., Wang, C., Tao, J., Zhang, X., Zhang, C. Y., and et.al,
Y. Z. (2023). Audio deepfake detection: A survey.
Zaman, K., Marchisio, A., Hanif, M. A., and et.al., M. S.
(2024). A survey on quantum machine learning: Cur-
rent trends, challenges, opportunities, and the road
ahead.
Zhang, Y., Wang, W., and et.al., P. Z. (2021). The effect of
silence and dual-band fusion in anti-spoofing system.
In Interspeech.
Hybrid Classical Quantum Learning Model Framework for Detection of Deepfake Audio
239