Received 11 February 2024, accepted 22 February 2024, date of publication 26 February 2024, date of current version 6 March 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3370838

Federated Deep Learning for Monkeypox Disease

Detection on GAN-Augmented Dataset
DIPANJALI KUNDU1,2 , MD. MAHBUBUR RAHMAN1 , ANICHUR RAHMAN2 , DIGANTA DAS3 ,
UMME RAIHAN SIDDIQI4 , MD. GOLAM RABIUL ALAM 5 , (Member, IEEE),
SAMRAT KUMAR DEY6 , GHULAM MUHAMMAD 7 , (Senior Member, IEEE),
AND ZULFIQAR ALI8 , (Member, IEEE)
1 Department of Computer Science and Engineering, Military Institute of Science and Technology, Dhaka 1216, Bangladesh
2 Department of Computer Science and Engineering, National Institute of Textile Engineering and Research (NITER), Constituent Institute of the University of
Dhaka, Savar, Dhaka 1350, Bangladesh
3 Department of Statistics, Jahangirnagar University, Dhaka 1342, Bangladesh
4 Physiology Department, Mymensingh Medical College, Mymensingh 2207, Bangladesh
5 Department of Computer Engineering, Brac University, Dhaka 1212, Bangladesh
6 School of Science and Technology (SST), Bangladesh Open University (BOU), Gazipur 1705, Bangladesh
7 Department of Computer Engineering, College of Computer and Information Sciences, King Saud University, Riyadh 11421, Saudi Arabia
8 School of Computer Science and Electronic Engineering, University of Essex, CO4 3SQ Colchester, U.K.

Corresponding authors: Ghulam Muhammad (ghulam@ksu.edu.sa), Anichur Rahman (anis_cse@niter.edu.bd), and Md. Mahbubur
Rahman (mahbub@cse.mist.ac.bd)
The authors acknowledge the Researchers Supporting Project number (RSP2024R34), King Saud University, Riyadh, Saudi Arabia.

ABSTRACT After the coronavirus disease 2019 (COVID-19) outbreak, the viral infection known as
monkeypox gained significant attention, and the World Health Organization (WHO) classified it as a global
public health emergency. Given the similarities between monkeypox and other pox viruses, conventional
classification methods encounter difficulties in accurately identifying the disease. Furthermore, sharing
sensitive medical data gives rise to concerns about security and privacy. Integrating deep neural networks
with federated learning (FL) presents a promising avenue for addressing the challenges of medical data
categorization. In light of this, we propose an FL-based framework using deep learning models to classify
monkeypox and other pox viruses securely. The proposed framework has three major components: (a) a
cycle-consistent generative adversarial network to augment data samples for training; (b) deep learning-based
models such as MobileNetV2, Vision Transformer (ViT), and ResNet50 for the classification; and (c) a
flower-federated learning environment for security. The experiments are performed using publicly available
datasets. In the experiments, the ViT-B32 model yields an impressive accuracy rate of 97.90%, emphasizing
the robustness of the proposed framework and its potential for secure and accurate categorization of
monkeypox disease.

INDEX TERMS Cycle GAN, deep neural network, federated learning, WHO, convolution neural network,
monkeypox detection, vision transformer, datasets, data analysis.

I. INTRODUCTION public health concern. Although its main impact is in Central

Monkeypox is an infectious disease originating from wildlife and West Africa, the disease has extended to urban areas
and affecting humans. Its symptoms closely resemble those and is prevalent near tropical rainforests. This global public
observed in smallpox patients. Since the eradication of small- health significance is not limited to African regions but
pox in 1980 [1] and the cessation of smallpox immunizations, extends worldwide, as evidenced by instances of monkeypox
monkeypox has emerged as the primary orthopoxvirus of appearing in non-endemic countries as recently as May
2022 [2]. The COVID-19 pandemic has underscored the
The associate editor coordinating the review of this manuscript and critical need to swiftly isolate confirmed viruses to contain
approving it for publication was Leimin Wang . their spread [3], [4]. Delaying the identification process not

2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 32819
 D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

only hampers treatment but also elevates transmission rates, TABLE 1. Some terminologies with their description in alphabetically
ordered.
potentially leading to pandemics. Thus, finding practical
ways to enhance early identification is imperative for curbing
the dissemination of this fatal ailment. Researchers have
turned to various machine learning (ML) techniques, partic-
ularly deep learning methods, for medical image analysis to
effectively differentiate the monkeypox virus from other pox
diseases [5].
Given that monkeypox disease shares symptomatic simi-
larities with other conditions like chickenpox and smallpox,
the differentiation of pox viruses solely based on symptoms
can be challenging. In this context, automated detection
emerges as a pivotal factor. Implementing automated detec-
tion methods through image processing and leveraging ML
or Deep Learning (DL) offers a viable solution [6], [7]. terminologies used in this work, with their descriptions in
However, the training and testing of DL models necessitate alphabetical order. Section II reviews related research work
a substantial volume of data. Hence, the early detection and identifies gaps in the current literature. Section III
of monkeypox disease using neural networks faces some describes the research methodologies and procedures used
challenges. One of the biggest challenges, in this case, is the for the study and explains the rationale behind their choice.
limitation of the dataset [8], [9]. As it is a very rare disease, After that, Section IV discusses experimental setup and
a lack of data might cause the neural network training process data collection, and provides experimental results. Further,
to overfit. In such a circumstance, models trained with a small Section V interprets the results in the context of the
dataset fail to perform adequately. research questions, and provides a discussion; and finally,
Generative Adversarial Networks (GAN) can be used to in Section VI, the key findings are summarised, the research
add more data, which is a useful way to deal with the question’s significance is reiterated, the contributions are
problems that come up when training deep neural network highlighted, and a final assessment with a recommendation
models with small or uneven datasets. This strategy aims is made.
to create a balanced dataset, thus enhancing the efficacy of
model construction. The purpose of this study is to develop II. LITERATURE REVIEW
a unique framework for the classification of monkeypox Recently, medical professionals have been utilizing tech-
virus lesions using medical image data that blends federated nological advancements to aid in disease identification.
learning (FL) with CycleGAN [10]. CycleGAN serves The healthcare sector is significantly advantaged by the
as a proficient tool for generating synthetic image data, integration of deep learning methodologies in computer
while federated learning is a decentralized machine learning vision for computer-aided diagnosis. Many researchers have
strategy that facilitates collaborative work across various suggested the integration of deep learning models with image
entities without exposing private data. Integrating deep neural analysis in the domain of skin diseases, particularly in the
networks that excel in image recognition and classification accurate detection of the pox virus within skin lesions.
tasks with the federated learning environment enhances Altun et al. [13] presented a Convolutional Neural Network
the classification approach’s efficiency. This combination (CNN) with transfer learning that achieved 96% efficiency.
not only ensures robust classification capabilities but also The models offered in this paper include MobileNetV3,
enforces additional security measures [11], [12]. Based on the ResNet50, VGG-19, DeneNet121, and Xception where
preceding context, the primary contributions of this study are MobileNetV3 outperformed the others. Pramanik et al. [14]
as follows: suggested a collaborative approach for the detection of
the monkeypox skin lesion with a five-fold evaluation
• We consider FL and the Vision Transformer to build a
procedure. Their approach received an accuracy score of 0.93.
secure framework for the categorization of Pox virus
Sitaula and Shahi [15] performed a comparative analysis
skin lesions. In particular, we use the Flower FL and
among thirteen different DL models for the classification
investigate several vision transformer models, such as
of monkeypox skin images. The best accuracy achieved
ViT-B16 and ViT-B32.
in this work is 87% with 85% precision and recall.
• We augment the Pox virus image dataset by incorporat-
To explain and ensure the transparency of the predicted
ing the Cycle GAN data augmentation method.
model, Azar et al. [5] proposed several methods like LIME
• The proposed framework achieves higher accuracy in
and Grad-CAM for model explainability with a DenseNet
the FL-DL environment with the vision transformer
model for the classification task and achieved an accuracy
model, ViT-B32, and with augmented data.
of 97%. Again, Ahsan et al. [16] proposed six models with
The rest of the paper is organized systematically, with the explainability for comparative analysis of the performance
following major sections: Table 1 represents the technical of different models for the classification of monkeypox.

32820 VOLUME 12, 2024

D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

As a related study shows, Jaradat et al. [17] used a centers or hospitals is still challenging as these images
comparative analysis of different models, such as VGG-16, contain sensitive information about a patient. Due to security
EfficientNetB3, and others, to find the best and most efficient concerns, patients sometimes show less interest in sharing
model for classifying monkeypox skin diseases. MobileNet data with third parties.
got the best results, with an accuracy of almost 98%. In the literature, the majority of research on detecting
However, the work discussed above used a dataset that is monkeypox is confined to the utilization of traditional
from an open-source repository [18] that includes 228 images deep-learning techniques. Some authors have advocated for
of monkeypox skin lesions and other Pox virus skin sores. ensemble learning [14], [15] to maximize the accuracy of
Because the images of the monkeypox skin lesions are their developed models, and certain methods have been
limited, additional measures are necessary for the training carefully adapted for specific tasks [5], [16]. Most papers
of the models for disease classification. Ahsan et al. [19] have employed standard data augmentation techniques due
implemented generalization and regularization-based transfer to the scarcity of training data [13], [27]. However, standard
learning models for monkeypox detection and showed that image augmentation techniques have limitations, as they may
their proposed optimized ResNet-101 can achieve the best not cover all possible variations that a model might encounter
performance for multiclass classification with an accuracy of in real-world scenarios. Moreover, advanced augmentation
99%. Saleh and Rabie [20] proposed the Human Monkeypox techniques, such as GANs have been explored to generate
Detection (HMD) system for monkeypox detection using more realistic and diverse augmented data. However, the
a blood test dataset. They introduced an Improved Binary existing literature only involves a few applications of
Chimp Optimization (IBCO) algorithm for feature selection GAN [23], [26]. Only a few papers have addressed privacy
in the ensemble model designed for monkeypox detection. concerns related to data [28]. To our knowledge, there is
Their proposed HMD strategy achieved an accuracy of no existing research that has introduced methodologies for
98.48% with 350 training data samples. Privacy is a major detecting monkeypox using federated learning.
concern in medical diagnosis. Blockchain technology can In summary, this article has proposed a secure framework
ensure the privacy and integrity of user’s data. In [21], in response to the challenges mentioned earlier. This
Gupta et al. proposed a blockchain-embedded monkeypox framework eliminates the necessity for data sharing among
detection and classification framework using transfer learn- parties by fusing FL with DL techniques. Given the limited
ing. Their proposed RssNet50 model achieves an accuracy of dataset on monkeypox skin lesions, we generated synthetic
98.80%. images using CycleGAN.
Researchers proposed several ways for synthetic image
generation, like applying GAN for data augmentation or III. PROPOSED METHODOLOGY
transforming the existing images. Among them, GAN is The proposed approach is divided into two parts. First,
an efficient process that has been widely used in synthetic we assessed our deep learning models independently, and
image data generation that uses ML and DL models. then we incorporated our deep learning models in a federated
Rashid et al. [22] analyzed several works and pointed out the learning environment to gauge their efficiency in both
challenges of the application of deep Neural Networks (NN) cases. Due to the limited dataset, our primary step involved
in the medical domain as the number of available datasets generating enhanced images using GAN. Following that,
is limited. The authors investigated the GAN technique for we employed a deep NN model for image classification.
synthetic image generation to create a robust model for the We then split the images into two batches for training. One
skin lesion classification task. Qin et al. [23] augmented skin batch received datasets for monkeypox and normal skin
images using style-based GAN and then trained and tested images, while the other was provided with datasets for other
the images using deep neural networks with transfer learning, diseases like measles and chickenpox. A detailed explanation
and this process improved the accuracy of the system by of our approach is presented in this section. In Fig. 1, the
1.6%. To overcome the shortcomings and imbalances of proposed system architecture is depicted.
melanoma skin cancer images, Zhao et al. [24] implemented
Style GAN for the generation of a balanced and vast amount A. DATA AUGMENTATION WITH CYCLEGAN
of data. Sedigh et al. [25] presented a CNN-based skin Without the need for paired training data, the Cycle
cancer detection framework that integrated GAN for the Consistent Generative Adversarial Network, or CycleGAN,
processing of synthetic images from the available image is a model developed using machine learning that creates
dataset. However, in [26], the author augmented data using synthetic images from sample data. Two NNs, a generator
GAN and showed that the augmented data can be the same and a discriminator, are used in this system to output images
as the real image data, but on the other hand, the hardware concurrently. The translation is performed bidirectionally,
and software specification required by the GAN-based model as suggested by the name of the architecture. Furthermore, the
can sometimes be challenging to implement in the medical discriminator’s role is to assess the excellence of the image
sector as this can impose an extra cost to the user. Although generated by the translation process between the domain
these methods are chosen for producing synthetic images, of healthy skin images and the domain of monkeypox skin
the collection of medical images across several medical images, and vice versa. This approach is suggested to be
VOLUME 12, 2024 32821
 D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

FIGURE 1. Proposed architecture with GAN and federated learning model.

TABLE 2. The hyperparameter settings of the proposed model. Four deep learning models: MobileNet, ResNet50, ViT-B32,
and ViT-B16, have been chosen for the federated learning
image categorization task. The classification task’s core is
based on these deep neural models. The server, or central
server, then initialized the global model. This global model
awaits client information. The clients connect to the global
utilized in this study to generate synthetic images so that server, then download the global model, and the local training
the framework can function with unpaired data. CycleGAN process using local datasets begins. The clients send their
is appropriate since the dataset for this field of monkeypox changes to the global model without their datasets. Once
detection or Pox virus classification is small and obtaining the central server has received all of the updates, it uses
the necessary images is challenging. Fig. 2 presents the the FedAvg method, which is shown in Equation 1 [32],
architecture of CycleGAN [29]. to combine them.

ci
B. DEEP LEARNING MODEL FOR IMAGE CLASSIFICATION g 1X
In this study, to address the challenge of training with the pt+1 = δi ∗ pit (1)
ci
limited dataset, we leverage a pre-trained deep neural network i=1

model combined with transfer learning, enhancing the

g
system’s performance. Through transfer learning, knowledge Here pt+1 is the global model update at the time (t+1),
from the extensive ImageNet dataset can be applied to our ci is the number of clients that take part in the averaging,
smaller, domain-specific dataset [30]. We have assessed both δi is the weights added to each client during the averaging
heavyweight and lightweight DL models to measure the process and pit is the local model parameter on device i at
models’ effectiveness. The pre-trained models we utilized time t.
include MobileNet, the vision transformers (ViT-B16, and
ViT-B32), and ResNet50 [31]. The description of the model’s 1) MOBILENET
hyperparameters is given in Table 2. MobileNet is a compact, deep artificial neural network
variation that is mostly utilized in embedded computers
C. FEDERATED LEARNING WITH DEEP LEARNING MODEL for image-related procedures. The depthwise convolutions
FOR IMAGE CLASSIFICATION enable this neural model’s neural network to have fewer
The federated learning environment used in this study is parameters, which lowers the cost of computation. The
implemented using the Flower federated learning framework. process of Mobilenet is first the depthwise convolution

32822 VOLUME 12, 2024

D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

FIGURE 2. CycleGAN architecture with generator and discriminator.

followed by the point-wise convolution [33].

Ci
X
T ∗K = Ti ∗ K (2)
i
In Equation 2, the convolution process is shown where T
is the tensor as input, and Ti and K ) which is the Kernal
represent the i-th element of the tensor T, the kernel or
the filter, respectively, and * represents the convolution
operation. After doing the element-wise product and sliding
the kernel over the input tensor in the convolutional layer, the
output of the convolution operation is found by adding the
two.

2) VISION TRANSFORMER MODEL (VIT-B16 AND VIT-B32)

The Transformer, which is the basis of ViT models, excels
in visual classification tasks that leverage self-attention.
This method allows the model to consistently focus on
diverse sections of image data. To capture the attributes
of the entire image, it is segmented into patches, which
are then fed to the encoder. For the transformer model
known as ViT-B32, an image patch size of 32 is utilized.
Conversely, the B-16 model comprises 16 layers and uses a
patch size of 16. Both the ViT-B32 and ViT-B16 transformer
versions are designed for image processing. ViT-B32 uses
larger patches (size 32), while ViT-B16 uses smaller patches
(size 16). Additionally, ViT-B16’s ability to capture local
and global properties may suffer from having 16 fewer
layers than ViT-B32. Both of these models represent basic FIGURE 3. Data distribution and sample images.
designs [34].

3) RESNET50 component of design that aids in resolving issues with

ResNet50 belongs to the residual network family and is training extremely deep networks, including the vanishing
a deep convolutional neural network. Because of its deep gradient issue. The model performs well in situations where
architecture, which consists of 50 layers, it can capture comprehending complicated feature hierarchies is essential,
complex patterns and characteristics in images. The word such as object identification and image classification, because
‘‘residual’’ refers to the utilization of residual blocks, a crucial of the depth of its design [35].

VOLUME 12, 2024 32823

 D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

respectively. For experiments, we split the dataset into

80 percent and 20 percent of the train and test datasets for
both cases (augmented and without augmentation). Further,
for validation purposes, we kept from the original dataset
(without augmentation) only 38 images of monkeypox,
10 images of chickenpox, 10 images of measles, and
29 normal skin images; these images were not used in
training or testing. In the case of augmentation, 110 samples
from 2914 images of monkeypox, 79 from 2640 images of
chickenpox, 65 from 2643 of measles, and 127 from 2826 of
normal were used for validation, and these images were not
used for training or testing.

B. PERFORMANCE INDICATORS
The accuracy score, recall, specificity, then F1-score, and
curve of ROC are the performance metrics for the assessment
of monkeypox identification using image categorization. The
values of true positive (TruePos), true negative (TrueNeg),
rate of false positive, and rate of false negative are used to
compute each of these metrics using the following equations.
True Pos
Precision(P) = (3)
True Pos + False Pos
True Pos
Recall(R) = (4)
FIGURE 4. Three sample of data augmentation through cycle GAN.
True Pos+ False Neg
2∗P∗R
F1S = (5)
P+R
IV. EXPERIMENTAL DESIGN AND ANALYSIS OF THE In 3 here P is referred to as Precision, in 4 R is referred to
RESULTS as Recall, and in 5 the F1S is refereed as F1 score.
A. DATASET DESCRIPTION
A dataset of skin lesions [36] is used in the experiments. C. EXPERIMENTAL ENVIRONMENT
We combine this dataset with a dataset from [37] to make In this section, we describe the environment where the
the classification of Pox viruses into four classes. Digital classification task was experimented. In this study, we used
color photographs of skins are the main images that were TPU over GPU among the several hardware accelerators
used in this research. Monkeypox, chickenpox, measles, and accessible in Colab since TPU is thought to be more
normal skin images are four types of skin lesions available in energy-efficient and also specifically made for tensors.
the collected data. The combined dataset consists of a total To make use of TPU, we use the Colab Pro version. Due to
of 381 images of monkeypox, 102 images of chickenpox, the use of both heavy and light models in this work as well as
110 images of measles, and 293 images of normal skin. the need for energy economy, we used TPU as the hardware
All the color images in the open-source collection were accelerator with High RAM in the Python 3 environment. The
specifically focused on capturing the afflicted portions of the proposed federated learning framework was implemented
patient’s skin. Since the digital input consists of RGB-colored using the Flower framework, proposed in [43].
images, each image contains three channels. The images
exhibit diverse formats and sizes due to their collection from D. ANALYSIS OF THE RESULT
many sources, hence making them inappropriate for use In this proposed method, four DL models MobileNetv2,
in predictive analytics. To address this problem, all of the ViT-B16, ViT-B32, and ResNet50 have been utilized for the
images have been resized to meet the exact requirements classification task. Table 3 shows the accuracies under the
set by the models. We have included images of normal FL environment, and for comparison purposes, it also shows
skin as well as images obtained from other data sources to the accuracies without the FL environment. In both cases,
create disease images. The CycleGAN is used to generate model accuracies when trained with and without augmented
augmented images. After applying CycleGAN to our dataset, datasets are also shown. We used four clients, each having
the dataset size increased significantly. In Fig. 3a, the image different datasets. The best accuracy obtained by the global
dataset distribution of each class for Original, Augmented, model is 97.90% by leveraging ViT-B32 under the secured
and Cumulative is shown. Fig. 3b shows some samples FL framework. Fig. 5 shows the performance matrix graphs
of image data. Fig. 4a, 4b, 4c represent the sample of of the ViT-B32 model. Fig. 5 (a) shows the ROC curve
augmented data images of monkeypox, normal, and measles, for four classes, obtained F1-Score, recall, and precision

32824 VOLUME 12, 2024

D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

FIGURE 5. Performance indicators of ViT-B32.

FIGURE 6. Comparative result analysis.

values are also depicted in Fig. 5 (b) and the corresponding As indicated in Fig. 6, the performance of the ViT model is
values are presented in Table 5, and the training vs validation notably poorer without data augmentation when compared to
accuracy, loss curve against epochs and confusion matrix the other deep learning models. Yet, with an increased num-
are shown in Fig. 5(c), 5 (d), and 5 (e), respectively. Again, ber of datasets, the ViT model showcases remarkable accu-
we have added a dropout layer and considered running racy improvements beyond the other models. Conversely,
the model up to 30 epochs with early stopping to the the RestNet 50 performs well in scenarios involving both
Vit-B32 model that reserved the validation loss, and this augmented and unaugmented datasets. Nevertheless, due to
model stopped after 30 epochs to ensure the model does not its computational complexity, the model might not always be
overfit. suitable, especially within interconnected systems featuring

VOLUME 12, 2024 32825

 D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

TABLE 3. Calculated system accuracy with and without FL environment.

TABLE 4. A comparative assessment of the proposed framework to various cutting-edge models.

TABLE 5. Obtained score of the three metrics. augmentation. This approach allows the classification of skin
images into either monkeypox or other conditions across
various clients, all the while maintaining the confidentiality
of image data. By adopting this method, the necessity for
data sharing is reduced, thereby addressing concerns related
to data breaches. Particularly within the medical industry,
sensor devices [39]. Based on the results, it becomes evident our suggested approach holds the potential to enhance
that the models specified exhibit superior performance when imaging-related tasks with greater precision. By confining
augmented with CycleGAN-generated data, as opposed to user data to the local network, we ensure data security [44].
utilizing an unaugmented dataset. However, it’s worth noting Throughout this study, deep learning models [45], including
that the ResNet50 model requires more time compared to MobileNetV2, Vision Transformers (ViT-B16 and B32),
the others, necessitating higher configuration memory for and ResNet50, were employed for training and calculating
computation. the global model’s parameters. Furthermore, a comparison
was drawn between the federated learning environment
V. DISCUSSION and the deep learning network model without security
This study proposes a monkeypox detection approach measures. The comparative outcomes reveal that the model’s
based on federated learning coupled with CycleGAN data accuracy remains consistent while user data confidentiality

32826 VOLUME 12, 2024

D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

is safeguarded within the federated learning environment. [2] J. Kaler, A. Hussain, G. Flores, S. Kheiri, and D. Desrosiers, ‘‘Monkeypox:
Again, a comparative analysis of existing work with the A comprehensive review of transmission, pathogenesis, and manifesta-
tion,’’ Cureus, vol. 14, no. 7, pp. 1–11, 2022.
model used and the accuracy achieved has been presented [3] A. Rahman, M. S. Hossain, G. Muhammad, D. Kundu, T. Debnath,
in Table 4. From the table, it is visible that our proposed M. Rahman, M. S. I. Khan, P. Tiwari, and S. S. Band, ‘‘Federated learning-
model can achieve higher accuracy while ensuring security. based AI approaches in smart healthcare: Concepts, taxonomies, chal-
lenges and open issues,’’ Cluster Comput., vol. 26, no. 4, pp. 2271–2311,
In extending the application to other predictive tasks in Aug. 2023.
healthcare, this proposed framework can be utilized to [4] A. Rahman, C. Chakraborty, A. Anwar, M. R. Karim, M. J. Islam,
safeguard patient data privacy and construct robust models D. Kundu, Z. Rahman, and S. S. Band, ‘‘SDN–IoT empowered intelligent
for critical illness prediction tasks. framework for industry 4.0 applications during COVID-19 pandemic,’’
Cluster Comput., vol. 25, no. 4, pp. 2351–2368, Aug. 2022.
To the best of our knowledge, the FL+DL and [5] A. Sorayaie Azar, A. Naemi, S. Babaei Rikan, J. Bagherzadeh Mohasefi,
FL+DL+CycleGAN approaches that have been suggested H. Pirnejad, and U. K. Wiil, ‘‘Monkeypox detection using deep neural
are the first to be used in recent years to identify monkeypox. networks,’’ BMC Infectious Diseases, vol. 23, no. 1, p. 438, Jun. 2023.
[6] H. Altaheri, G. Muhammad, and M. Alsulaiman, ‘‘Dynamic convolution
Furthermore, no study that we compared in this research with multilevel attention for EEG-based motor imagery decoding,’’ IEEE
addressed the application of a federated learning strategy. Internet Things J., vol. 10, no. 21, pp. 18579–18588, Jul. 2023.
Only the deep technique was compared in Table 4 to draw [7] F. Yasmin, Md. M. Hassan, M. Hasan, S. Zaman, C. Kaushal, W. El-Shafai,
comparisons between approaches that are comparable or and N. F. Soliman, ‘‘PoxNet22: A fine-tuned model for the classification
of monkeypox disease using transfer learning,’’ IEEE Access, vol. 11,
nearly equivalent and have previously been employed for the pp. 24053–24076, 2023.
diagnosis of monkeypox. Furthermore, when this comparison [8] F. Alshehri and G. Muhammad, ‘‘A few-shot learning-based ischemic
assessment was designed, the same number and model were stroke segmentation system using weighted MRI fusion,’’ Image Vis.
Comput., vol. 140, Dec. 2023, Art. no. 104865.
not taken into account. [9] S. Islam, U. Sara, A. Kawsar, A. Rahman, D. Kundu, D. Das,
A. N. M. Rezaul, and M. Hasan, ‘‘SGBBA: An efficient method for
VI. CONCLUSION prediction system in machine learning using imbalance dataset,’’ Int. J.
Adv. Comput. Sci. Appl., vol. 12, no. 3, pp. 1–12, 2021.
This paper introduces a secure Federated Learning and deep [10] T. Kwak and Y. Kim, ‘‘Semi-supervised land cover classification of remote
learning-based framework for monkeypox virus detection sensing imagery using CycleGAN and EfficientNet,’’ KSCE J. Civil Eng.,
using skin lesion images. The primary objective is to vol. 27, no. 4, pp. 1760–1773, Apr. 2023.
[11] H. Li, C. Li, J. Wang, A. Yang, Z. Ma, Z. Zhang, and D. Hua, ‘‘Review
enhance classification performance while maintaining the
on security of federated learning and its application in healthcare,’’ Future
confidentiality of medical images. By aggregating model Gener. Comput. Syst., vol. 144, pp. 271–290, Jul. 2023.
parameters during training, data privacy is ensured, as sensi- [12] A. Rahman, K. Hasan, D. Kundu, M. J. Islam, T. Debnath, S. S. Band,
tive information remains undisclosed. Given the limited data and N. Kumar, ‘‘On the ICN-IoT with federated learning integration of
communication: Concepts, security-privacy issues, applications, and future
availability, the CycleGAN generator plays a crucial role in perspectives,’’ Future Gener. Comput. Syst., vol. 138, pp. 61–88, Jan. 2023.
augmenting training and test data. The resulting synthetic [13] M. Altun, H. Gürüler, O. Özkaraca, F. Khan, J. Khan, and Y. Lee,
images are partitioned into four groups, each corresponding ‘‘Monkeypox detection using CNN with transfer learning,’’ Sensors,
vol. 23, no. 4, p. 1783, Feb. 2023.
to a client, and utilized to train local models using classifiers
[14] R. Pramanik, B. Banerjee, G. Efimenko, D. Kaplun, and R. Sarkar,
such as MobileNet, ViT-B16, ViT-B32, and ResNet50. ‘‘Monkeypox detection from skin lesion images using an amalgamation
Among these, ViT-B32 excels with an impressive accuracy of CNN models aided with beta function-based normalization scheme,’’
of 97.90. To validate the effectiveness of the proposed deep PLoS ONE, vol. 18, no. 4, Apr. 2023, Art. no. e0281815.
[15] C. Sitaula and T. B. Shahi, ‘‘Monkeypox virus detection using pre-trained
learning-based Federated Learning approach, the DL models deep learning-based approaches,’’ J. Med. Syst., vol. 46, no. 11, p. 78,
are also assessed individually outside the FL environment, Oct. 2022.
and they perform nearly as well. The adoption of Federated [16] M. M. Ahsan, M. S. Ali, M. M. Hassan, T. A. Abdullah,
K. D. Gupta, U. Bagci, C. Kaushal, and N. F. Soliman, ‘‘Monkeypox
Learning not only ensures user data privacy within the diagnosis with interpretable deep learning,’’ IEEE Access, vol. 11,
local network but also accomplishes the categorization task pp. 81965–81980, 2023.
effectively. This approach is especially relevant in medical [17] A. S. Jaradat, R. E. Al Mamlook, N. Almakayeel, N. Alharbe, A. S. Almu-
contexts where limited datasets and data privacy concerns flih, A. Nasayreh, H. Gharaibeh, M. Gharaibeh, A. Gharaibeh, and H.
Bzizi, ‘‘Automated monkeypox skin lesion detection using deep learning
are prevalent. In the future, further security measures, such and transfer learning techniques,’’ Int. J. Environ. Res. Public Health,
as Blockchain, could be incorporated to enhance the security vol. 20, no. 5, p. 4422, Mar. 2023.
of the skin lesion images. [18] S. Nafisa Ali, M. Tazuddin Ahmed, T. Jahan, J. Paul, S. M. Sakeef Sani,
N. Noor, A. Nahar Asma, and T. Hasan, ‘‘A web-based mpox skin lesion
Authors Contributions All authors contributed to this detection system using state-of-the-art deep learning models considering
research’s design, analysis, writing, and revising. All authors racial diversity,’’ 2023, arXiv:2306.14169.
approved the submitted version of the manuscript. [19] M. M. Ahsan, M. R. Uddin, M. S. Ali, M. K. Islam, M. Farjana, A. N. Sakib,
K. A. Momin, and S. A. Luna, ‘‘Deep transfer learning approaches for
Conflict of Interest The authors declare no competing monkeypox disease diagnosis,’’ Exp. Syst. Appl., vol. 216, Apr. 2023,
financial and non-financial interests. Art. no. 119483.
[20] A. I. Saleh and A. H. Rabie, ‘‘Human monkeypox diagnose (HMD)
strategy based on data mining and artificial intelligence techniques,’’
REFERENCES Comput. Biol. Med., vol. 152, Jan. 2023, Art. no. 106383.
[1] J. G. Breman and D. A. Henderson, ‘‘Diagnosis and management of [21] A. Gupta, M. Bhagat, and V. Jain, ‘‘Blockchain-enabled healthcare
smallpox,’’ New England J. Med., vol. 346, no. 17, pp. 1300–1308, monitoring system for early monkeypox detection,’’ J. Supercomput.,
Apr. 2002. vol. 79, no. 14, pp. 15675–15699, Sep. 2023.

VOLUME 12, 2024 32827

 D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

[22] H. Rashid, M. A. Tanveer, and H. Aqeel Khan, ‘‘Skin lesion [42] A. H. Alharbi, S. K. Towfek, A. A. Abdelhamid, A. Ibrahim, M. M. Eid,
classification using GAN based data augmentation,’’ in Proc. 41st D. S. Khafaga, N. Khodadadi, L. Abualigah, and M. Saber, ‘‘Diagnosis
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2019, of monkeypox disease using transfer learning and binary advanced dipper
pp. 916–919. throated optimization algorithm,’’ Biomimetics, vol. 8, no. 3, p. 313,
[23] Z. Qin, Z. Liu, P. Zhu, and Y. Xue, ‘‘A GAN-based image synthesis method Jul. 2023.
for skin lesion classification,’’ Comput. Methods Programs Biomed., [43] K. H. Li, P. P. B. de Gusmão, D. J. Beutel, and N. D. Lane, ‘‘Secure
vol. 195, Oct. 2020, Art. no. 105568. aggregation for federated learning in flower,’’ in Proc. 2nd ACM Int.
[24] C. Zhao, R. Shuai, L. Ma, W. Liu, D. Hu, and M. Wu, ‘‘Dermoscopy image Workshop Distrib. Mach. Learn., Dec. 2021, pp. 8–14.
classification based on StyleGAN and DenseNet201,’’ IEEE Access, vol. 9, [44] F. Alshehri and G. Muhammad, ‘‘A comprehensive survey of the Internet
pp. 8659–8679, 2021. of Things (IoT) and AI-based smart healthcare,’’ IEEE Access, vol. 9,
[25] P. Sedigh, R. Sadeghian, and M. T. Masouleh, ‘‘Generating synthetic pp. 3660–3678, 2021.
medical images by using GAN to improve CNN performance in skin cancer [45] M. S. I. Khan, A. Rahman, T. Debnath, M. R. Karim, M. K. Nasir,
classification,’’ in Proc. 7th Int. Conf. Robot. Mechatronics (ICRoM), S. S. Band, A. Mosavi, and I. Dehzangi, ‘‘Accurate brain tumor detection
Nov. 2019, pp. 497–502. using deep convolutional neural network,’’ Comput. Struct. Biotechnol. J.,
[26] A. Bissoto, E. Valle, and S. Avila, ‘‘GAN-based data augmentation vol. 20, pp. 4733–4745, 2022.
and anonymization for skin-lesion analysis: A critical review,’’ in Proc.
IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW),
Jun. 2021, pp. 1847–1856.
[27] D. Bala, M. S. Hossain, M. A. Hossain, M. I. Abdullah, M. M. Rahman,
B. Manavalan, N. Gu, M. S. Islam, and Z. Huang, ‘‘MonkeyNet: DIPANJALI KUNDU received the B.Sc. degree
A robust deep convolutional neural network for monkeypox disease in computer science and engineering from
detection and classification,’’ Neural Netw., vol. 161, pp. 757–775, Chittagong University of Engineering and Tech-
Apr. 2023. nology (CUET), Bangladesh, in 2018. She is
[28] A. Rahman, J. Islam, D. Kundu, R. Karim, Z. Rahman, S. S. Band, currently pursuing the M.Sc. degree in CSE with
M. Sookhak, P. Tiwari, and N. Kumar, ‘‘Impacts of blockchain in software- the Military Institute of Science and Technology
defined Internet of Things ecosystem with network function virtualization (MIST). She has been a Lecturer in computer
for smart applications: Present perspectives and future directions,’’ Int. J. science and engineering (CSE) with the National
Commun. Syst., p. e5429, Feb. 2023. Institute of Textile Engineering and Research
[29] H. Wang, Q. Qi, W. Sun, X. Li, B. Dong, and C. Yao, ‘‘Classification (NITER), Savar, Dhaka, Bangladesh, since 2020.
of skin lesions with generative adversarial networks and improved Her research interests include machine learning, human–computer inter-
MobileNetV2,’’ Int. J. Imag. Syst. Technol., vol. 33, no. 5, pp. 1561–1576, action, the Internet of Things (IoT), blockchain (BC), software defined
Sep. 2023. networking (SDN) 5G, industry 4.0, and robotics.
[30] S. I. Khan, A. Shahrior, R. Karim, M. Hasan, and A. Rahman, ‘‘MultiNet:
A deep neural network approach for detecting breast cancer through multi-
scale feature fusion,’’ J. King Saud Univ. Comput. Inf. Sci., vol. 34, no. 8,
pp. 6217–6228, Sep. 2022.
[31] E. Hassan, M. S. Hossain, A. Saber, S. Elmougy, A. Ghoneim, and MD. MAHBUBUR RAHMAN received the
G. Muhammad, ‘‘A quantum convolutional network and ResNet Ph.D. degree in information computer science
(50)-based classification architecture for the MNIST medical from Japan Advanced Institute of Science and
dataset,’’ Biomed. Signal Process. Control, vol. 87, Jan. 2024, Technology (JAIST), Japan, in 2004. He is cur-
Art. no. 105560. rently a Running Professor with the Department
[32] M. M. Rahman, D. Kundu, S. A. Suha, U. R. Siddiqi, and S. K. of Computer Science and Engineering (CSE),
Dey, ‘‘Hospital patient’s length of stay prediction: A federated learning Military Institute of Science and Technology
approach,’’ J. King Saud Univ.-Comput. Inf. Sci., vol. 34, no. 10, (MIST), Bangladesh. He has authored more than
pp. 7874–7884, 2022. 75 research articles in reputed journals and con-
[33] R. K. Shukla and A. K. Tiwari, ‘‘Masked face recognition using MobileNet ferences globally. His research interests include
v2 with transfer learning,’’ Comput. Syst. Sci. Eng., vol. 45, no. 1,
computer networks, image processing, pattern recognition, bioinformatics,
pp. 293–309, 2023.
and machine learning.
[34] H. Zheng, G. Wang, and X. Li, ‘‘Identifying strawberry appearance quality
by vision transformers and support vector machine,’’ J. Food Process Eng.,
vol. 45, no. 10, Oct. 2022, Art. no. e14132.
[35] I. Z. Mukti and D. Biswas, ‘‘Transfer learning based plant diseases
detection using ResNet50,’’ in Proc. 4th Int. Conf. Electr. Inf. Commun. ANICHUR RAHMAN received the B.Sc. and
Technol. (EICT), Dec. 2019, pp. 1–6. M.Sc. degrees in computer science and engineer-
[36] S. N. Ali, M. T. Ahmed, J. Paul, T. Jahan, S. M. S. Sani, N. Noor, and ing from Mawlana Bhashani Science and Technol-
T. Hasan, ‘‘Monkeypox skin lesion detection using deep learning models: ogy University (MBSTU), Tangail, Bangladesh,
A feasibility study,’’ 2022, arXiv:2207.03342. in 2017 and 2020, respectively. He has been
[37] D. Bala. (2022). Monkeypox Skin Images Dataset (MSID). [Online]. an Assistant Professor in computer science and
Available: https://www.kaggle.com/dsv/3971903 engineering (CSE) with the National Institute
[38] M. Lakshmi and R. Das, ‘‘Classification of monkeypox images using of Textile Engineering and Research (NITER),
LIME-enabled investigation of deep convolutional neural network,’’
Constituent Institute of the University of Dhaka,
Diagnostics, vol. 13, no. 9, p. 1639, May 2023.
Savar, Dhaka, Bangladesh, since 2020. He has
[39] T. Nayak, K. Chadaga, N. Sampathila, H. Mayrose, N. Gokulkrishnan,
authored more than 30 articles in high-ranking journals and conferences,
M. Bairy, G. S. Prabhu, and S. Umakanth, ‘‘Deep learning based detection
of monkeypox virus using skin lesion images,’’ Med. Novel Technol. including FGCS, IEEE INTERNET OF THINGS JOURNAL, IEEE AC, DCAN, JISA,
Devices, vol. 18, Jun. 2023, Art. no. 100243. SR, Nature, CSBJ, and CCJ. His research interests include the Internet
[40] A. K. Gairola and V. Kumar, ‘‘Monkeypox disease diagnosis using machine of Things (IoT), blockchain, software defined networking (SDN), network
learning approach,’’ in Proc. 8th Int. Conf. Signal Process. Commun. function virtualization (NFV), AI, image processing, machine learning, 5G,
(ICSC), Dec. 2022, pp. 423–427. industry 4.0, and data science. He has received the Best Paper Award at
[41] S. Örenç, E. Acar, and M. S. Özerdem, ‘‘Utilizing the ensemble of deep the International Conference on Trends in Computational and Cognitive
learning approaches to identify monkeypox disease,’’ Dicle Üniversitesi Engineering (TCCE 2020). His articles received the best paper award and
Mühendislik Fakültesi Mühendislik Dergisi, vol. 13, no. 4, pp. 685–691, nominations at different international conferences. He is also a reviewer of
2022. high-quality journals and conferences.

32828 VOLUME 12, 2024

D. Kundu et al.: Federated Deep Learning for Monkeypox Disease Detection on GAN-Augmented Dataset

DIGANTA DAS received the B.Sc. degree in GHULAM MUHAMMAD (Senior Member,
computer science and engineering from the Chit- IEEE) received the B.S. degree in computer
tagong University of Engineering and Technology science and engineering from Bangladesh Uni-
(CUET), Bangladesg, in 2018, and the M.Sc. versity of Engineering and Technology, in 1997,
degree in data science from Jahangirnagar Uni- and the M.S. and Ph.D. degrees in electronic
versity, in 2023. His research interests include and information engineering from Toyohashi
machine learning, data science, business intelli- University and Technology, Japan, in 2003 and
gence, fuzzy logic systems, and robotics. 2006, respectively. He is currently a Professor with
the Department of Computer Engineering, College
of Computer and Information Sciences, King Saud
University (KSU), Riyadh, Saudi Arabia. He was a recipient of the Japan
Society for Promotion and Science (JSPS) fellowship from the Ministry of
Education, Culture, Sports, Science and Technology, Japan. His research
UMME RAIHAN SIDDIQI received the M.B.B.S. interests include signal processing, machine learning, the IoT, medical signal
degree from the Sher-E-Bangla Medical College and image analysis, AI, and biometrics. He has authored or coauthored more
(SBMC), Bangladesh, in 2003, and the Doctor than 300 publications, including IEEE/ACM/Springer/Elsevier journals and
of Medicine (M.D.) degree in physiology from flagship conference papers. He owns two U.S. patents. He received the
Bangabandhu Sheikh Mujib Medical University Best Faculty Award from the Computer Engineering Department, KSU,
(BSMMU), Bangladesh, in 2019. She has been from 2014 to 2015. He has supervised more than 15 Ph.D.’s and master’s
an Assistant Professor with Mymensingh Medical theses. He is involved in many research projects as a principal investigator
College (MMC), Bangladesh, since April 2023. and a co-principal investigator.
She has authored more than 20 research articles
in reputed journals and conference proceedings.
Her current research interests include autism spectrum disorder in children,
neurophysiology, health informatics, and AI.

MD. GOLAM RABIUL ALAM (Member, IEEE)

received the B.S. degree in computer science and
engineering, the M.S. degree in information tech-
nology, and the Ph.D. degree in computer engi-
neering from Kyung Hee University, South Korea,
in 2001, 2011, and 2017, respectively. He was
a Postdoctoral Researcher with the Computer
Science and Engineering Department, Kyung Hee
University, from March 2017 to February 2018.
He is currently a Full Professor with the Depart-
ment of Computer Science and Engineering, Brac University, Bangladesh.
His research interests include healthcare informatics, mobile cloud and edge
computing, ambient intelligence, and persuasive technology. He is a member
of the IEEE IES, CES, CS, SPS, CIS, KIISE, and IEEE ComSoc. He received
several best paper awards at prestigious conferences.
ZULFIQAR ALI (Member, IEEE) received the
M.Sc. and M.S. degrees in computer science from
the University of Engineering and Technology
Lahore, Pakistan, in 2007 and 2010, respectively,
SAMRAT KUMAR DEY received the B.Sc. and the Ph.D. degree in electrical and electronic
degree in computer science and engineering from engineering from Universiti Teknologi Petronas,
Patuakhali Science and Technology University Malaysia, in 2017. He was a Researcher with the
(PSTU), Bangladesh, in 2015, and the M.Sc. Department of Computer Engineering, King Saud
degree in computer science and engineering from University, from 2010 to 2018, and a Research
the Military Institute of Science and Technology Fellow with the BT Ireland Innovation Center,
(MIST), BUP, Bangladesh, in 2019. He is currently Ulster University, from 2018 to 2020. He is currently a Lecturer with the
a Lecturer with the School of Science and Tech- School of Computer Science and Electrical Engineering, University of Essex,
nology, Bangladesh Open University. Previously, Colchester, U.K. He has published more than 60 international peer-reviewed
he was an Assistant Professor with the Department conference papers and journal articles. His current research interests include
of Computer Science and Engineering (CSE), Dhaka International University explainable AI, digital speech and image processing, privacy and security in
(DIU), Bangladesh. His research interests include visual data analytics, healthcare using watermarking, and audio forgery detection. He is a fellow
health informatics, information visualization, human–computer interaction, of the Higher Education Academy and Advance HE, U.K. He has served on
machine learning, and deep learning. He served as an Academic Editor for the technical program committees for the IEEE Smart World Congress and
PLOS One and Computational and Mathematical Methods (Wiley-Hindawi). IEEE ACAI. He is also serving as an Associate Editor for the IEEE JOURNAL
He has also regularly served as a technical committee member and a reviewer OF BIOMEDICAL AND HEALTH INFORMATICS.
for different international journals and conferences worldwide.

VOLUME 12, 2024 32829