Skin Cancer Classification and Detection Using Federated Learning

Malliga Subramanian, Kalaivani B, Jeevasree G, Mathan Kumar A and Nandhini P S

Department of Computer Science and Engineering, Kongu Engineering College, Erode, Tamil Nadu, India

Keywords: Skin Cancer Classification, Federated Learning, CNN, MobileNetV2, Federated Averaging

Abstract: Detecting skin cancer involves challenges like ensuring the secure and private handling of sensitive medical

data. Traditionally, centralized models have been used for classification and diagnosis, but these can risk data

leaks and compromise patient privacy. To address this, a distributed learning system is proposed, allowing

data to remain private while maintaining model accuracy. In this paper, we introduce a federated learning

model for skin cancer classification. This system uses four independent clients: two trained on the ISIC 2018

dataset (with 7 skin disease types) and two trained on the ISIC 2019 dataset (with 8 disease types). The weights

from the clients are combined and updated using the FedAvg algorithm to create a global model without

sharing raw data between clients. The clients use CNN and MobileNetV2 for building the classifiers. This

federated learning approach not only ensures data privacy but also achieves better performance, surpassing

the current state-of-the-art accuracy for skin cancer classification across different datasets

1 INTRODUCTION

Early and accurate diagnosis of skin cancer in the

dermatology field is the most critical factor in

improving survival rates. Conventionally,

dermatologists have always relied on medical

history and physical examination that depends on

direct visual diagnoses of cutaneous lesions to

determine the presence of skin cancers. However,

this form of diagnosis is prone to inherent human

mistakes in the process, as well as subjective

judgment when time is of the essence. The

integration of machine learning and artificial

intelligence has led to the emergence of automated

systems that aid in the diagnosis of skin diseases.

However, centralization of sensitive medical data

towards training the machine learning model does

pose a privacy concern regarding sensitive

information in health care where security for data is

concerned.

Federated learning(FL) addresses privacy

concerns by enabling multiple institutions or devices

to collaboratively train a model without sharing

raw patient data, ensuring data privacy and security

while leveraging distributed data sources. In this

approach, each institution, or client, trains a model

locally on its own dataset and shares just the learned

model updates (the changes in the weight) with the

central server. This has allowed learning while

keeping sensitive patient data private and compliant

with regulations like HIPAA and GDPR.

The proposed work introduces a federated

learning- based system that utilizes the architecture

of Convolutional Neural Networks and

MobileNetV2 to classify skin diseases and cancer. In

this work, four local clients are used for training.

The ISIC 2018 dataset is utilized by two of the

clients that covers seven kinds of diseases, while the

remaining two clients utilize the ISIC 2019 dataset

which comprises nine varieties of diseases. By using

transfer learning, wherein pre-trained models are

fine-tuned on local datasets, the system can achieve

high classification accuracy without requesting large

amounts of labelled data. Once trained, each client

submits its model updates to a central server, which

aggregates them using the FedAvg algorithm to

create a global model.

The rest of the article is structured as follows: In

section 2, we review the research attempts related to

the skin disease classification. Section 3 explains the

proposed methodology and the set of experiments

conducted. The results of the experiments are

presented in Section 4 along with a discussion and

finally, we provide the conclusion of our work in

Section 5.

Subramanian, M., B, K., G, J., A, M. K. and P S, N.

Skin Cancer Classiﬁcation and Detection Using Federated Learning.

DOI: 10.5220/0013586200004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 31-40

ISBN: 978-989-758-763-4

31

2 LITERATURE REVIEW

Below we present an overview of the recent attempts

to classify the skin cancer using deep learning

models.

(Sandler, Howard et al. 2018) used

MobileNetV2, a highly iterative version of the

original MobileNet to combat some major issues

with deep learning on resource-constrained devices

such as smartphones and embedded systems. A

balance between being computationally efficient and

giving an enhanced feature extraction performance

has made the model useful in image classification,

object detection, or even medical diagnosis

applications.

(Guan, Yap et al. 2024) reviews the FL methods-

those methods that allow the collaborative training

of machine learning models without sharing the

sensitive medical data. The authors classify the FL

approaches into three major categories: client-side

learning, server-side aggregation, and

communication optimization. In addition to

presenting empirical experiments on the FL

performance for medical imaging, the authors

highlight challenges, benchmark datasets, and

software platforms.

The work by (Hossen, Panneerselvam et al.,

2022) titled "Decentralized Training of a Model for

Skin Disease Classification based on FL, while

ensuring Data Privacy and Security issues related to

Internet of Medical Things" describes skin disease

classification with the help of FL in the

decentralized model and CNN. It aims at dealing

with data privacy and security issues associated with

the Internet of Medical Things.

(Ali, Shaikh et al. 2022) investigates automated

classification of several skin cancer types by

EfficientNet architectures. The main goal is to

enhance early diagnosis and prevention by

improving reliable, deep learning-based diagnostic

tools. This study can be said to take a step towards

advanced dermatology by implementing efficient AI

models in practice.

(Gautam et al., 2024) explores the

implementation of deep learning techniques,

particularly convolutional neural networks (CNNs),

for the detection of skin cancer. The main focus of

this research is to analyze dermoscopic images for

improving diagnostic accuracy, in addition to the

early detection of skin cancer, to develop automated

tools for dermatological diagnosis.

(Lilhore et al. 2024) presented an accurate skin

cancer diagnosis model based on the combination of

a hybrid U-Net and an enhanced MobileNet-V3

architecture using techniques that perform

hyperparameter optimization. The results were an

enhanced performance in segmentation and

classification of skin lesions.

(Agbley et al. 2021) investigated multimodal

melanoma detection using federated learning in

enhancing privacy and collaboration between

different datasets. The methodology fused various

data modalities, such as images and patient

metadata, in a way that enhances diagnostic

precision. Their study underscored the potential of

federated learning to support safe and efficient

training without the need to centralize sensitive data.

From the review of the recent attempts, we

understand that despite the challenges related to non-

IID data are discussed, yet tailored solutions for

diverse dermatological datasets across demographics

remain underexplored. Scalability and deployment

in real-world, resource-constrained healthcare

facilities, along with the impact of communication

costs in FL, are insufficiently addressed.

3 IMPLEMENTATION

3.1 Federated Learning

Federated learning enables secure, collaborative

model training across decentralized datasets from

various hospitals, clinics, and personal devices. In

conventional machine learning, all data are

centralized at one server; however, with federated

learning the patient data stay on local devices,

detailing only trained model parameters or weights

with a central server. Hence, this provides privacy

and compliance with the regulations regarding

medical applications, such as HIPAA and GDPR.

Local models are trained using pretrained CNN

and MobileNet-V2, which are excellent at

extracting features and making classifications. After

training on their own dermatoscopic datasets, the

updated model weights are sent to a central server.

The server combines these weights from different

models using a method called Federated Averaging

(FedAvg), creating a global model that works well

across multiple datasets. This global model is then

sent back to local devices for further improvement in

the next training rounds.

This approach makes the models more reliable

and reduces errors caused by relying on a single

dataset. FL allows for accurate and robust skin

cancer detection while keeping data private, making

it adaptable for use in hospitals or personal devices

INCOFT 2025 - International Conference on Futuristic Technology

32

Figure 1: Federated Learning Architecture

of any size. Figure 1 depicts the proposed FL

architecture.

3.2 Federated Averaging

The foundation of FL is an algorithm called

Federated Averaging (FedAvg), which makes it

possible to create machine learning models that are

trained across different clients using their own

private data. In our work, we use FedAvg to develop

the global model. The crux of FedAvg lies in getting

clients to train their local models on their data and

then send the model updates, i.e., weights, to the

server instead of sending raw data, thus keeping

sensitive information within the client. The average

weight updates are done as in Equation (1).

𝑊=

1

𝑁

𝑛



𝜔







(1)

Where,

•

𝜔

𝑖

is the model weights for clients 𝑖.

•

𝑛

𝑖

is the amount of data for clients 𝑖.

•

𝑘 is the total number of clients.

•

𝑁 is the total number of data points across all

clients.

•

W is the global model weight after

aggregation.

The process starts with each client training a

local model on their own data for several rounds.

Once the training is complete, the clients send their

model weights to a central server. The server

combines these weights to create a new global model,

which is then shared back with the clients. The clients

use this updated global model as a starting point for

the next round of local training.

This cycle repeats until the model reaches the

best possible performance. The FedAvg method

reduces communication costs by sharing weights

instead of gradients, making it efficient for low-

bandwidth situations and ideal for applications that

require strong privacy.

3.3 Experiments

To demonstrate the purpose and performance of FL,

we conduct the following experiments.

3.3.1 First Experiment

In first experiment, we train local models based on

CNN on 100% of the ISIC 2018 training data. To

Skin Cancer Classiﬁcation and Detection Using Federated Learning

33

test the generalizing capability of the local models,

we run the model on ISCI 2019 dataset.

3.3.2 Second Experiment

In second set of experiments, we use MobielNetV2

for training the local models. These models are

trained on 100% of the ISIC 2018 training data.

MobileNetV2 is efficient in computation footprint

besides retaining good accuracy. The generalizing

capability of the models are tested using ISIC 2019

dataset.

3.3.3 Third Experiment

Four clients are locally trained with the CNN and

aggregated globally using FedAvg. In this third

experiment, FL, integrated over four local clients,

led to two clients trained by ISIC 2018 dataset while

the remaining two were trained with ISIC 2019

dataset, to solve the overfitting phenomena observed

in the first two experiments. Common diseases were

taken and trained. After each local training, the

weights of the models were forwarded for global

averaging using the Federated Averaging Algorithm.

The further updated weights were then sent back to

each client.

Thus, the federated learning approach enables

the models to be trained using multiple datasets

without training the models on each dataset.

Although Client 1 trained on the ISIC 2018 dataset,

the updated global model could be used for new

images from the ISIC 2019 dataset and vice versa.

Hence, the models are generalized well concerning

previously unseen data by accumulating learning

evolved from both datasets into the federated

averaging process.

3.3.4 Fourth Experiment

Here, four clients are locally trained with the

MobileNetV2 and aggregated globally using

FedAvg. The fourth experiment was basically the

same as the third, except that the CNN model was

substituted for local training with MobileNetV2.

MobileNetV2 is built to give high accuracy while

being light on the number of parameters and

computations. Four local clients were set up: two

clients trained the ISIC 2018 dataset and the other

two the ISIC 2019 dataset. Similar to the previous

experiments, each local model was independently

trained on its respective datasets. Subsequently, the

model weights were sent to a central server for

federated averaging using the Federated Averaging

Algorithm. This algorithm averaged the weights

from the local models and updated them, after which

the global model weights were sent back to each

client, thus enabling the clients to benefit from the

knowledge learned by other clients.

4 RESULTS AND DISCUSSION

4.1 Local Models

In Experiment 1, a CNN model is trained and tested

on the ISIC 2018 datasets and achieved 83%

accuracy. This accuracy drastically fell to 62% when

applied the models on the ISIC 2019 datasets, which

means that the model has less generalizability. As

the model is trained by 2018. It is not suitable for

2019 images.

4.1.1 Classification results for ISIC 2018

using CNN

Table 1 shows the performance metrics of the

developed models for each class. CNN achieved an

accuracy of 83% on the test dataset of 938 samples.

CNN exhibited a good performance among the

classes with an F1-score of 0.93 for class 5. Nearly

39000 images was trained by CNN model as a single

client.

Table 1: Classification report for ISIC 2018 using CNN

Class Precision Recall F1-

Score

Support

0 0.53 0.35 0.42 26

1 0.58 0.50 0.54 30

2 0.41 0.37 0.39 75

3 0.13 0.33 0.19 6

4 0.34 0.51 0.41 39

5 0.94 0.92 0.93 751

6 0.61 1.00 0.76 11

Accuracy 0.83 938

Macro Avg 0.50 0.57 0.52 938

Weighted

Avg

0.84 0.83 0.83 938

Figure 2 visually represents the classification

performance of the CNN model on the ISIC 2018

dataset, highlighting the model's strengths and

weaknesses across different classes.

INCOFT 2025 - International Conference on Futuristic Technology

34

Figure 2: Confusion Matrix for ISIC 2018 using CNN

4.1.2 Classification results for ISIC 2019

using CNN

In Table 2, performance metrics of CNN is shown

for all the classes. The model gives an accuracy of

62% to the dataset of 5,074 samples. Of all the

classes, Class 4 showed the best performance, with

an F1-score of 0.79. Since the model is trained using

ISIC 2018 dataset, it gives the minimal performance

while testing with the 2019 dataset. The model is not

well suited for other datasets rather than the trained

dataset. Thus, it shows a limited capability of the

model in adapting to new forms of data distributions.

Table 2: Classification report for ISIC 2019 using CNN

Class Precision Recall F1-

Score

Support

0

0.31 0.14 0.19 175

1 0.49 0.59 0.53 665

2

0.33 0.28 0.30 526

3 0.60 0.06 0.11 49

4

0.76 0.82 0.79 2576

5 0.50 0.50 0.50 905

6 0.28 0.06 0.10 127

7 0.61 0.37 0.46 51

Accuracy 0.62 5074

Macro Avg 0.48 0.35 0.37 5074

Weighted

Avg

0.60 0.62 0.60 5074

The classification performance of the CNN

model on the ISIC 2019 dataset is shown in Figure

3.

Figure 3: Confusion Matrix for ISIC 2019 using CNN

Experiment 2 showed a slight improvement

using the MobileNetV2 model, with the ISIC 2018

dataset around 89% accuracy. MobileNetV2 proved

to have better trade-off between precision and recall

in all the classes, reducing the false positives and

negatives. Despite being highly accurate on the ISIC

2018 dataset, it fell short with its generalization

ability and large accuracy at 72% when tested on the

ISIC 2019 dataset. This might not be much as it was

better than the performance of the CNN model on

ISIC 2019, but it still points to some difficulty in

generalizing across datasets with very different

distributions.

4.1.3 Classification results for ISIC 2018

using MobileNetV2

Table 3 provides an overview of model performance

over all classes. The accuracy obtained is 89% for

test dataset. Apart from class 5, which gives an F1-

score of 0.96, making it the most prominent class in

the dataset, others stand out due to their excellent

performance on F1 scores: class 6 with 0.91 and

class 1 with 0.71. Since the CNN model is not

generalized, MobileNetV2 is used. This model gives

the better accuracy in testing the 2018 dataset but

while testing with 2019, the accuracy is not much

defined. But it has some improvement compared to

CNN model’s performance.

Skin Cancer Classiﬁcation and Detection Using Federated Learning

35

Table 3. Classification report for ISIC 2018 using

MobileNetV2

Class Precision Recall

F1-

Score

Support

0

0.80 0.15 0.26 26

1 0.60 0.87 0.71 30

2

0.76 0.56 0.65 75

3 0.60 0.50 0.55

6

4

0.41 0.62 0.49 39

5 0.96 0.97 0.96 751

6

0.91 0.91 0.91 11

Accurac

y

0.89 938

Macro

Av

g

0.72 0.65 0.65 938

Weighted

Avg

0.90 0.89 0.89 938

Classification performance of the MobileNetV2

model for ISIC 2018 is shown in Figure 4. The

findings indicate that yet powerful architecture of

MobileNetV2 would really fit skin cancer

classification tasks provided that training and testing

are performed on similar datasets.

Figure 4: Confusion Matrix for ISIC 2018 using

MobileNetV2

4.1.4 Classification results for ISIC 2019

using MobileNetV2

Table 4 gives the performance of MobileNetV2 on a

dataset comprising 5,074 samples, with the overall

accuracy of 72%. Class 4 was the best class

performance with an F1 score of 0.86. Figure 5

shows the performance of MobileNetV2 on ISIC

2019 dataset.

Table 4: Classification report for ISIC 2019 using

MobileNetV2

Class Precision Recall

F1

Support

0

0.50 0.18 0.27 175

1 0.58 0.86 0.69 665

2

0.63 0.36 0.46 526

3 0.50 0.02 0.04 49

4

0.83 0.88 0.86 2576

5 0.64 0.58 0.61 905

6

0.32 0.23 0.26 127

7 0.47 0.65 0.55 51

Accurac

y

0.72 5074

Macro Av

g

0.56 0.47 0.47 5074

Weighted

Avg

0.71 0.72 0.70 5074

Figure 5: Confusion Matrix for ISIC 2019 using

MobileNetV2

4.2 Federated Models

To combat the generalization problem, federated

learning CNN was used in Experiment 3, for

instance. All four local clients, trained on ISIC 2018

and ISIC 2019 datasets, have achieved 82% for ISIC

2018 and 76% for ISIC 2019 overall with better

generalization across datasets.

4.2.1 Classification results for ISIC 2018

using Fed-CNN

From Table 5, we can see that the performance of

CNN and it gives 82% accuracy. Four clients trained

their local CNN models using their own data and

shared their updated weights with a central server.

The server aggregated these weights by calculating

their weighted average and redistributed the updated

INCOFT 2025 - International Conference on Futuristic Technology

36

weights back to the clients. This iterative process

allowed the global model to benefit from diverse

client datasets while preserving data privacy. The

global model achieved 82% accuracy on ISIC 2018

and 76% on ISIC 2019, demonstrating improved

generalization compared to standalone training.

Table 5: Classification report for ISIC 2018 using Fed-

CNN

Class Precision Recall

F1-

Score

Support

0

0.14 0.02 0.04 40

1 0.70 0.44 0.54 153

2

0.68 0.70 0.68 124

3 1.00 0.28 0.44 20

4

0.35 0.51 0.42 185

5 0.60 0.63 0.61 385

6 0.51 0.43 0.47 31

Accuracy 0.82 938

Macro

Avg

0.57 0.43 0.46 938

Weighted

Avg

0.56 0.54 0.54 938

Figure 6 shows that the model performs best for

nevus category, with 243 correct predictions, but

also shows some misclassifications.

Figure 6: Confusion Matrix for ISIC 2019 using Fed-CNN

4.2.2 Classification results for ISIC 2019

using Fed-CNN

Next, we present the performance of CNN using

ISIC 2019 dataset with FL in Table 6. Compared to

the standalone CNN model trained only on ISIC

2018 data (Experiment 1), this approach

demonstrated improved performance on the ISIC

2019 dataset. The iterative process of weight

aggregation and redistribution enhanced the ability

of the models to generalize to datasets from different

distributions. This aggregation step allowed the

global model to benefit from the knowledge

acquired by all clients, effectively combining their

individual learning outcomes. This is visually

represented in Figure 7

Table 6: Classification report for ISIC 2019 using Fed-

CNN

Class Precision Recall

F1-

Score

Support

0

0.74 0.50 0.59 172

1 0.62 0.89 0.73 619

2

0.63 0.44 0.51 419

3 0.50 0.01 0.03 75

4

0.84 0.90 0.87 2798

5 0.73 0.58 0.65 869

6

0.30 0.23 0.26 122

Accuracy 0.76 5074

Macro

Av

g

0.60 0.52 0.52 5074

Weighted

Av

g

0.75 0.76 0.74 5074

Figure 7: Confusion Matrix for ISIC 2019 using Fed-

CNN

.

4.2.3 Classification results for ISIC 2018

using Fed-MobileNetV2

Experiment 4 has shown that the Federated

Averaging (FedAvg) method can be used with

MobileNetV2 as a local model architecture. Each

Skin Cancer Classiﬁcation and Detection Using Federated Learning

37

client was training its local MobileNetV2 model on

its data and sending the updated weights to a central

server.

The server, in turn, weighted these updated

weights for aggregation before sending the

distributed updated weights back to the clients. This

collaborative approach can iterate, where the global

model learns diverse data without privacy loss.

The MobileNetV2 Global model has been able to

achieve an accuracy of 80% in ISIC 2018 and 87%

in ISIC 2019. This can be referenced in table 7.

During these tests, the study achieved significant

improvements in generalization-performance. The

results were particularly significant for ISIC 2019.

The improvements show the power of combining the

FedAvg algorithm and using MobileNetV2

architecture in distributed skin cancer classification

tasks.

Table 7: Classification report for ISIC 2018 using Fed -

MobileNetV2

Class Precision Recall

F1-

Score

Support

0 0.14 0.02 0.04 21

1 0.04 0.04 0.04 52

2

0.26 0.41 0.32 49

3 1.00 0.39 0.56 8

4

0.82 0.85 0.84 534

5 0.56 0.51 0.53 259

6

0.51 0.48 0.49 15

Accuracy 0.80 938

Macro

Avg

0.48 0.39 0.40 938

Weighted

Av

g

0.66 0.66 0.66 938

Figure 8: Confusion Matrix for ISIC 2018 using Fed-

MobileNetV2

It is shown in Figure 8 that MobileNetV2

performed best for class 4 with 453 correct

predictions, demonstrating that it has a very good

identification ability for that particular class.

4.2.4 Classification results for ISIC 2019

using Fed-MobileNetV2

As shown in Table 8, Class 5 achieves an excellent

performance with its F1-score of 0.97 and with good

precision and recall, denoting that the class is clearly

the most represented class by a strong classification

of itself. Thus, the model has an overall accuracy of

87%, justifying its high competence in classifying

most cases correctly.

Table 8: Classification report for ISIC 2019 using Fed -

MobileNetV2

Class Precision Recall

F1-

Score

Support

0 0.16 0.14 0.15 122

1 0.60 0.67 0.63 170

2 0.64 0.52 0.57 354

3 0.23 0.43 0.30 32

4 0.36 0.50 0.42 209

5 0.97 0.97 0.97 4043

6 0.59 0.30 0.40 144

Accurac

y

0.87 5074

Macro

Avg

0.51 0.50 0.49 5074

Weighted

Av

g

0.87 0.87 0.87 5074

Figure 9: Confusion Matrix for ISIC 2019 using Fed-

MobileNetV2

INCOFT 2025 - International Conference on Futuristic Technology

38

Figure 9 shows that class 5 has the highest correct

predictions, with 2708 samples accurately classified.

This shows that the model performs exceptionally

well for this class. Class 2 also performs well, with

164 correct predictions, indicating good

discrimination for this class.

4.3 Discussion

The CNN model built under Experiment 1 with the

ISIC 2018 dataset achieved an accuracy of 83%

when deployed as a test on the same dataset. The

model's application on the ISIC 2019 dataset,

recorded an accuracy of only 62%. It indicates the

limited ability of the model to generalize the

assigned task in data from different distributions

such as ISIC 2019. Next, the MobileNetV2 model

was trained on the ISIC 2018 dataset, and it

performed better, achieving a high accuracy of 89%

when tested on the test dataset. When tested using

the ISIC 2019 database, this model showed a higher

level of generalization than the CNN, at 72%.

However, the performance gap shows that the model

is not entirely prepared for classifying images from

the 2019 database with training focusing only on the

2018 data. FL method aggregated model weights of

different clients and redistributed the new weights to

each client for refining the local models training. It

proved to be much more effective in building the

models that generalizes well to a wider variety of

datasets other than the ISIC 2018 dataset.

By applying this methodology, we noticed the

improved performance in both models. The local

CNN model with global FedAVG’S performance for

the classification of 2019 images increased to 76%

from 72% whereas for 2018 images the performance

slighted reduced to 82%. However, the model is well

suited for the different datasets. To have an

efficiency in classification, MobileNetV2 was used

as a local client model under the FedAvg algorithm.

This configuration achieved its highest accuracy,

with an 80% for ISIC 2018 and an impressive score

of 87% for the ISIC 2019 dataset. This model is well

suited for both datasets since the model weights are

aggregated globally and the updated weights are sent

back to all the four clients.

5 CONCLUSION AND FUTURE

WORK

In this work, we have designed a skin cancer

classification system using the federated learning

approach with the MobileNetV2 model. This

method has the potential to perform efficient,

privacy-preserving training on decentralized devices

while maintaining high performance. The final

accuracy of the model was 87% for ISIC 2019

dataset and 80% for ISIC 2018 dataset, making it a

promising candidate for use in real-world

applications such as skin cancer detection.

MobileNetV2 is a lightweight architecture and, thus,

more suitable for an edge device with a balance

between performance and computational efficiency.

We ensured that there was no sensitive medical

information leaving the user's device by using

federated learning to preserve privacy. Future work

would involve improving the dataset by applying

SMOTE , model optimization along with federated

learning methods used, and would allow continuous

learning, which makes the system faster on real-time

skin cancer detection on mobile devices. Federated

learning techniques could also be extended to make

this process better in terms of efficiency as well as

accuracy while training models. Such techniques as

Federated Averaging could be combined with even

more advanced techniques such as differential

privacy or secure multi-party computation to further

enhance the privacy and security of the model

updates.

REFERENCES

Sandler, M., A. Howard, M. Zhu, A. Zhmoginov and L.C.

Chen., 2018. Mobilenetv2: Inverted residuals and

linear bottlenecks. Proceedings of the IEEE conference

on computer vision and pattern recognition

Guan, H., P.-T. Yap, A. Bozoki and M. Liu., 2024.

"Federated learning for medical image analysis: A

survey." Pattern Recognition: 110424

Hossen, M. N., V. Panneerselvam, D. Koundal, K. Ahmed,

F. M. Bui and S. M. Ibrahim., 2022. "Federated

machine learning for detection of skin diseases and

enhancement of internet of medical things (IoMT)

security." IEEE journal of biomedical and health

informatics 27(2): 835-841.

Ali, K., Z. A. Shaikh, A. A. Khan and A. A. Laghari.,

2022. "Multiclass skin cancer classification using

EfficientNets–a first step towards preventing skin

cancer." Neuroscience Informatics 2(4): 100034.

Gautam, G. K., S. Singh and A. Singh., 2024. Skin Cancer

Identification Using Deep Learning Technique. 2024

International Conference on Communication,

Computer Sciences and Engineering (IC3SE), IEEE.

Lilhore, U. K., S. Simaiya, Y. K. Sharma, K. S. Kaswan,

K. B. Rao, V. M. Rao, A. Baliyan, A. Bijalwan and R.

Alroobaea., 2024. "A precise model for skin cancer

diagnosis using hybrid U-Net and improved

Skin Cancer Classiﬁcation and Detection Using Federated Learning

39

MobileNet- V3 with hyperparameters optimization."

Agbley, B. L. Y., J. Li, A. U. Haq, E. K. Bankas, S. Ahmad,

I. O. Agyemang, D. Kulevome, W. D. Ndiaye, B.

Cobbinah and S. Latipova., 2021. Multimodal

melanoma detection with federated learning. 2021

18th international computer conference on wavelet

active media technology and information processing

(ICCWAMTIP), IEEE.

Zakariah, M., M. Al-Razgan and T. Alfakih., 2024. "Skin

cancer detection with MobileNet-based transfer

learning and MixNets for enhanced diagnosis." Neural

Computing and Applications: 1-31.

Sumaiya, N. and A. Ali., 2024. "Federated Learning

Assisted Deep learning methods fostered Skin Cancer

Detection: A Survey." Frontiers in Biomedical

Technologies.

Ahmad, G., M. Saleem, J. A. Malik, W. A. Bukhari, M. I.

Kashif, H. Salahuddin, M. A. U. Rehman and A. U.

Rehman., 2024. "Mobile Application for Skin Disease

Classification Using CNN with User Privacy." Journal

of Computing & Biomedical Informatics.

INCOFT 2025 - International Conference on Futuristic Technology

40