Enhanced Credit Card Fraud Detection Using Federated Learning,
LSTM Models, and the SMOTE Technique
Weddou Mohamedhen
1
, Maha Charfeddine
2 a
and Yessine Hadj Kacem
1 b
1
CES Laboratory, National Engineering School of Sfax, University of Sfax, Sfax, Tunisia
2
REsearch Groups in Intelligent Machines, National Engineering School of Sfax (ENIS), Sfax, Tunisia
Keywords:
Credit Card Fraud, Federated Learning, Class Imbalance, SMOTE, LSTM, CNN, ADAM, Data Privacy.
Abstract:
In recent years, credit card transaction fraud has caused significant financial losses for both consumers and
financial institutions. To effectively combat these losses, the development of a sophisticated fraud detection
system is necessary. However, credit card fraud detection (CCFD) presents significant challenges, particularly
in regards to data security and privacy, limiting financial institutions’ ability to share transaction data for model
training. This paper introduces the use of Federated Learning for CCFD, a technique that allows for decen-
tralized learning while protecting data privacy. Federated Learning enables multiple institutions to collaborate
on model training without having to share sensitive data, effectively addressing privacy concerns. To address
the problem of class imbalance in fraud detection datasets, we apply the Synthetic Minority Oversampling
Technique (SMOTE) to ensure a balanced dataset. Our study compares Long Short-Term Memory (LSTM)
networks to Convolutional Neural Networks (CNN) within a Federated Learning framework. The experimen-
tal results demonstrate that combining SMOTE and LSTM in a Federated Learning setup produces superior
performance. These findings highlight the strength of LSTM models in processing sequential transaction data
and reveal that Federated Learning, when paired with resampling techniques, strengthens fraud detection ac-
curacy.
1 INTRODUCTION
The prevalence of credit card transactions has surged
due to the expansion of e-commerce, electronic bank-
ing, and mobile payments, leading to a significant in-
crease in credit card fraud. This rise has resulted in
global fraud losses reaching $33 billion in 2022, pro-
jected to reach $43 billion by 2026 (Report, 2023),
(Payments, 2023), (Moneyzine, 2023).
Fraudulent transactions are often carried out
through stolen or falsified credit card information,
making fraud detection a critical issue. Despite the
success of Machine Learning (ML) models, chal-
lenges such as dataset insufficiency and class imbal-
ance persist.
Dataset Insufficiency. Privacy concerns limit the
availability of public datasets, hindering the develop-
ment of robust fraud detection systems. To address
this, we use Federated Learning (FL), which enables
collaborative model development across institutions
a
https://orcid.org/0000-0003-2996-4113
b
https://orcid.org/0000-0002-5757-6516
while preserving privacy (Salam et al., 2023), (Zhang
and et al., 2021).
Skewed Class Distribution. Fraud detection datasets
are typically imbalanced, with fraudulent transactions
being rare. To overcome this, we apply the Synthetic
Minority Oversampling Technique (SMOTE) to bal-
ance the dataset (G. Bejjanki and Narsimha, 2018).
This paper proposes a novel approach combining
Federated Learning and SMOTE to tackle dataset in-
sufficiency and class imbalance. We optimize key FL
parameters, such as the fraction of participating insti-
tutions (F) and the number of local epochs (E), and
compare LSTM and CNN models. Our results show
that LSTM outperforms CNN in fraud detection. This
work aims to improve the accuracy and reliability of
fraud detection systems in financial institutions.
The paper is structured as follows: Section 2 re-
views existing approaches, Section 3 presents our
methodology, Section 4 outlines results, and Section
5 concludes with future work.
368
Mohamedhen, W., Charfeddine, M. and Hadj Kacem, Y.
Enhanced Credit Card Fraud Detection Using Federated Learning, LSTM Models, and the SMOTE Technique.
DOI: 10.5220/0013135100003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 368-375
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

2 RELATED WORK
Fraud detection algorithms in credit card transac-
tions employ Machine Learning techniques to ef-
fectively identify fraudulent activities. Traditional
approaches predominantly use centralized learning
models such as Decision Trees (DT), Random Forests
(RF), Logistic Regression (LR), Support Vector Ma-
chines (SVM), Extreme Gradient Boosting (XGB),
and unsupervised methods like Generative Adversar-
ial Networks (GAN), Auto-Encoders (AE), Restricted
Boltzmann Machines (RBM), and One-Class SVM
(OCSVM) (G. Bejjanki and Narsimha, 2018), (X. Niu
and Yang, 2019). While effective, these methods
face scalability and data privacy challenges due to
their centralized nature. For example, a study us-
ing the UCI Credit Card Dataset, comprising 284,807
transactions, highlighted the imbalanced nature of the
data with only 492 fraudulent transactions, impacting
model accuracy.
Federated Learning (FL) has emerged as a promis-
ing alternative, addressing concerns over data security
and privacy inherent in centralized models (M. Fahmi
and Nagati, 2016), (K. Chen and Zhang, 2019). FL
enables collaborative model training on distributed
data sources without sharing sensitive information,
thereby enhancing privacy while improving model
performance. Recent studies, such as one involving a
federated dataset of over 1 million transactions from
multiple banks, have demonstrated FL’s effectiveness
in real-time credit card fraud detection, showing sig-
nificant improvements compared to traditional cen-
tralized models (K. Chen and Zhang, 2019).
In Deep Learning, unsupervised models like AE
and RBM have shown high accuracy rates (88% to
94%) in detecting credit card fraud when integrated
into Federated Learning frameworks (al., 2019), (Su-
varna and Kowshalya, 2020). For instance, using the
IEEE-CIS Fraud Detection dataset, with over 590,000
records, these models required careful parameter tun-
ing and computational resources but offered robust
performance in identifying complex fraud patterns.
Privacy-preserving strategies such as combining FL
with differential privacy and homomorphic encryp-
tion further bolster the security of fraud detection sys-
tems (Albertio, 2019). (al., 2020).
Hybrid techniques integrating Decision Trees,
clustering algorithms, pairwise matching, Neural Net-
works, and genetic algorithms are also being explored
to predict fraud in various transactional datasets
citeb15, (Dornadula and Geetha, 2019). For exam-
ple, a hybrid approach on the European cardholders
dataset, consisting of 284,807 records, leveraged local
and global model characteristics to optimize perfor-
mance while minimizing communication overhead.
To address class imbalance challenges in fraud
detection datasets, various methods including cost-
sensitive Deep Learning approaches and resampling
techniques like SMOTE, EUS-Bag, and PSOAANN
have been developed (Kamaruddin and Ravi, 2016).
These techniques aim to balance dataset distribu-
tions and enhance model robustness against rare fraud
cases. A study using the Kaggle Credit Card Fraud
Detection dataset, which includes 284,807 transac-
tions, found that applying SMOTE improved the de-
tection rate of fraudulent transactions significantly,
though implementation requires careful consideration
to avoid biases and maintain computational efficiency.
Recent works have further advanced fraud detec-
tion in the FL framework. Salam et al. (2023) pro-
posed a Federated Learning model for credit card
fraud detection, incorporating data balancing tech-
niques to address privacy and class imbalance. Sim-
ilarly, Li and Walsh (2024) introduced FedGAT-
DCNN, combining Graph Attention Networks and di-
lated convolutions in FL to enhance fraud detection.
Our work builds upon these studies by integrating
SMOTE directly into FL and optimizing key FL pa-
rameters, such as the fraction of participating institu-
tions (F) and the number of local epochs (E), to en-
hance model performance and scalability.
Compared to existing approaches, our work offers
several significant value additions:
• Privacy Protection Through Federated Learn-
ing. Unlike traditional approaches that require
sharing raw data, our method allows secure col-
laboration among financial institutions, address-
ing data privacy concerns.
• Enhanced Fraud Detection with SMOTE. By
using SMOTE to balance the data, our ap-
proach overcomes the challenge of class imbal-
ance, thereby improving the performance of pre-
dictive models.
• Comparative Analysis of LSTM and CNN
Models. Our study provides a detailed compara-
tive analysis of LSTM and CNN networks within
a Federated Learning framework, offering valu-
able insights into the effectiveness of these mod-
els for credit card fraud detection.
• Practical Applicability for Financial Institu-
tions. Experimental results demonstrate that the
combination of SMOTE and LSTM within the
Federated Learning system yields the best results,
highlighting the superiority of the LSTM model
in handling sequential transaction data.
In summary, the combination of centralized and
Federated Learning models, along with advanced
Enhanced Credit Card Fraud Detection Using Federated Learning, LSTM Models, and the SMOTE Technique
369

Figure 1: Federated Learning architecture-based intelligent Credit Card Fraud Detection (CCFD) system.
Deep Learning techniques, privacy-preserving strate-
gies, and effective class imbalance management,
holds great promise for improving the reliability and
efficiency of credit card fraud detection systems. Fur-
ther research should look into these methodologies,
particularly their adaptability to changing fraud pat-
terns, in order to improve system security and perfor-
mance.
3 METHODOLOGY
Figure 1 depicts our proposed Intelligent Credit Card
Fraud Detection (CCFD) system built on a secure
Federated Learning architecture. The system begins
with input transaction and label data, followed by
preprocessing steps like cleaning, normalization, and
feature engineering (Ali et al., 2024a). Through our
Federated Learning framework, local models train
on distributed datasets without sharing raw data, and
an aggregation server combines them into a global
model. The model is evaluated using metrics like pre-
cision, recall, accuracy, and F1-score, and predicts the
likelihood of new transactions being fraudulent or le-
gitimate. The following sub-sections provide details
about our suggested architecture:
3.1 Dataset Description
The dataset used in this research was obtained from
Kaggle (Group—ULB, 2018). It consists of 284,807
anonymous credit card transactions from European
cardholders in September 2013, of which 492 are
fraudulent, resulting in a heavily skewed dataset.
The dataset includes 30 features that describe each
transaction, including transaction amount and time.
Clients are fully aware of all features, ensuring ef-
fective alignment of their data for federated learning.
There are no missing data points, maintaining the in-
tegrity of the analysis.
Table 1: Overview of the Dataset Obtained from
(Group—ULB, 2018).
Total Dataset #Fraud #Not Fraud Label Not F Label F
Dataset 284,807 492 284,315 0 1
3.2 Resampling Technique
Synthetic minority oversampling technique (SMOTE)
SMOTE is an effective oversampling method to ad-
dress class imbalance by generating synthetic exam-
ples of the minority class rather than duplicating ex-
isting ones. SMOTE selects similar instances from
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
370

the minority class and interpolates between them to
create new, synthetic data points, resulting in a more
balanced dataset and improved model performance as
depicted in Figure 2. Several studies have demon-
strated the efficacy of SMOTE in improving classi-
fier performance in imbalanced datasets, concerning
cyberattack detection challenges (Ali et al., 2024b).
Thus, to handle class imbalance, SMOTE is used to
create synthetic fraudulent examples, applied only to
the training set (XTrain).
Figure 2: Visualization of the Synthetic Minority Over-
sampling Technique (SMOTE) applied to balance the
dataset.
To reduce the content while maintaining the key
information, here’s a more concise version of the sec-
tion. This version retains the core explanations about
LSTM, CNN, and optimizers, focusing on the essen-
tials to fit within an 8-page limit:
3.3 Modeling
We developed two Deep Learning (DL) models,
Long Short-Term Memory (LSTM) and Convolu-
tional Neural Networks (CNN), to detect credit card
fraud. While Federated Learning ensures data pri-
vacy, it does not fully address the need for advanced
models. Integrating LSTM and CNN enhances the
ability to handle complex patterns in the data.
Long Short-Term Memory Networks (LSTMs).
(Li and Zhao, 2021) are designed to capture temporal
dependencies in sequential data, enabling the detec-
tion of fraud by modeling transaction sequences over
time.
Convolutional Neural Networks (CNNs). (Zhang
and Liu, 2022) are used here to capture local feature
patterns in transaction data, even though CNNs are
typically used for grid-like data. While GRU or RNNs
may be better suited for sequential data, CNNs are
included to explore their potential in fraud detection.
Optimizers play a key role in model performance:
ADAM (Adaptive Moment Estimation). (V. Felbab
and Horv
´
ath, 2021) adapts learning rates and speeds
up convergence, even with noisy data.
SGD (Stochastic Gradient Descent). (Akbar and
Chowanda, 2022) updates parameters using small
batches, helping to avoid local minima and train ef-
ficiently on large datasets.
By combining LSTM and CNN models within
Federated Learning and using advanced optimizers,
we improve fraud detection accuracy and reliability.
3.4 Federated Fraud Detection
Framework
In our Federated Learning-based fraud detection sys-
tem, C banks each hold a private dataset, and SMOTE
is used to address class imbalance by generating syn-
thetic minority class examples. The banks collaborate
to build a fraud detection model while maintaining
privacy, agreeing on a common model architecture,
including activation and loss functions, before train-
ing begins.
min
w∈R
d
1
n
n
∑
i=1
L
c
(x, y;w) (1)
where
L
c
(x, y;w) =
1
|D
c
|
∑
i∈D
c
L
c
(x
ci
, y
ic
;w) (2)
The server initializes the fraud detection model
parameters. During each communication round, a
random fraction of banks is selected to participate.
These banks download the current global model, com-
pute gradients on their private datasets, update their
local models, and send the updates back to the server.
The server then aggregates the updates to improve the
global model.
w
t+1
← w
t
− η
C
∑
c=1
1
n
∑
i∈D
c
∇L
c
(x
ci
, y
ic
;w) (3)
Considering the impact of skewed data on model
performance, we use the combination of data size and
detection model performance α
t+1
on each bank as
the weight of the parameter vector. It can be written
as:
w
t+1
← w
t
−
C
∑
c=1
n
c
∑
C
c=1
n
c
α
t+1
f
c
(4)
In Federated Learning (FL), each bank performs a
gradient descent step using its own data, and the
server aggregates these models via a weighted aver-
age, updating the global model over T iterations. FL
addresses privacy concerns by allowing collaborative
training on distributed data while maintaining confi-
dentiality. However, FL faces challenges like commu-
nication costs, which depend on factors such as the
fraction of participating banks, minibatch size, and
the number of local epochs. These parameters must
be optimized to balance parallelism and efficiency.
The training process is shown in Algorithm 1 and Fig-
ure 3.
Enhanced Credit Card Fraud Detection Using Federated Learning, LSTM Models, and the SMOTE Technique
371

Algorithm 1. FFD Framework.
Require: The private dataset of banks and financial
institutions
Ensure: A credit card fraud detection model with
Federated Learning
1: Server Update:
2: Initialize the detection classifier and its parame-
ters w
0
3: for each round t = 1, 2, . . . , T do
4: Randomly choose max(F ·C, 1) banks as N
t
5: for each bank c ∈ N
t
in parallel do
6: w
t+1
, α
t+1
← BankUpdate(n, w
t
)
7: end for
8: w
t+1
←
∑
T
t=1
C
n
C
α
t
+ 1
9: end for
10:
11: BankUpdate(n, w):
12: Data Processing: Rebalance raw dataset with
SMOTE (applied only to the training set) and
split it into two parts: 80% training data and 20%
testing data
13: Training:
14: B ← Split D
n
into batches of size B
15: for each local epoch i from 1 to E do
16: for each batch b ∈ B do
17: w ← w − η∇L(x, y;w)
18: end for
19: end for
20: Testing:
21: Return w and validation accuracy α to server
Figure 3: Federated Learning model for FDS.
4 EXPERIMENTAL RESULTS
This section compares the performance of LSTM and
CNN models in a secure Federated Learning frame-
work enhanced with SMOTE for credit card fraud de-
tection. We evaluated models based on precision, re-
call, F1-score, and accuracy, with hyperparameters in-
cluding a 30-input size, 128 hidden units across three
layers, and a learning rate of 0.001. ADAM and SGD
optimizers were tested over 20 epochs with varying
batch sizes. The setup involved 12 clients, with 10
clients per round over 100 communication rounds us-
ing TensorFlow Federated. The results are compared
to previous related works.
4.1 Performance Metrics
To evaluate credit card fraud detection systems, we
use several metrics beyond accuracy, particularly in
imbalanced datasets. While accuracy measures over-
all correctness, it can be misleading if fraudulent
transactions are misclassified. Therefore, we con-
sider precision, which reflects the system’s reliability
in identifying fraud, recall, which measures its effi-
ciency in detecting all fraudulent transactions, and the
F1 score, which is the harmonic mean of precision
and recall. These metrics provide a more compre-
hensive evaluation of the system’s performance (Ja-
vatpoint, 2024).
Accuracy =
T P + T N
T P + FP + T N + FN
(5)
Recall =
T P
T P + FN
(6)
F1 =
2 × Precision × Recall
Precision + Recall
(7)
Precision =
T P
T P + FP
(8)
4.2 Results and Discussions
In this section, performance metrics, including preci-
sion, recall, accuracy, and F1-score, have been dis-
cussed to ensure the effectiveness of all classifiers
used in conjunction with the chosen resampling tech-
nique, SMOTE. For model evaluation, the LSTM and
CNN algorithms were adopted for comparison. The
experiments were conducted using an 80:20 training-
to-testing ratio, which showed that LSTM outper-
forms CNN in federated learning for fraud detection
due to its ability to capture temporal dependencies,
whereas CNN is better suited for detecting local pat-
terns, which are less relevant for the sequential nature
of fraud data.
The Figure 4’ diagrams depict the performance
of the developped LSTM model with federated data
learning (SMOTE+ LSTM+ FDL), employing two
optimization techniques, notably ADAM and SGD,
across different batch size configurations and client
counts (numclients). Evaluation metrics include ac-
curacy , precision, recall , and F1 score. The results
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
372

illustrate how these metrics vary based on model con-
figuration parameters and the platform used. Each
experimental configuration was conducted with 12
clients and 10 selected for training, ensuring a fair
and representative comparison of performance. We
observe significant advancements in model perfor-
mance across these various metrics, highlighting the
efficacy of tailored approaches like (SMOTE + LSTM
+ FDL) in achieving high accuracy, precision, recall,
and overall balanced performance. The findings show
ADAM’s accuracy consistently of 0.999 value, while
SGD slightly around 0.998. Precision shows signif-
icant variability, reaching up to 0.887 with ADAM
and 0.646 with SGD, depending on the batch size.
Recall remains steady, ranging from 0.907 to 0.926,
and F1 Score fluctuates with values between 0.842
and 0.879. Regarding batch sizes, 256 seems to pro-
vide an optimal balance between precision and re-
call for the (SMOTE + LSTM + FDL) model, espe-
cially when using 100 iterations with the Adam op-
timizer and a learning rate of 0.001. These findings
underscore the importance of optimizer selection in
developing robust Deep Learning-based fraud detec-
tion systems, directly influencing model precision and
stability of predictions. The analysis also highlights
batch size’s crucial role in model performance, partic-
ularly. In conclusion, our study advocates for ADAM
as the preferred optimizer to maximize efficiency and
reliability of the combinaison (SMOTE + LSTM +
FDL)-based fraud detection systems.
Moreover, Figure 5 illustrates the performance
metrics of the developped CNN model using the two
different optimizers, SGD and ADAM, across varying
batch sizes (32, 64, 128, and 256) and consistent pa-
rameters of num clients = 12 and num selected = 10.
The results reveal that Accuracy remains consistently
high and stable, ranging from 0.998 to 0.999 across
all batch sizes and optimizers. Precision exhibits no-
ticeable variability, ranging from 0.470 to 0.880 for
SGD and 0.830 to 0.887 for ADAM, reflecting the
impact of optimizer choice on model precision. Re-
call shows relatively stable performance, hovering be-
tween 0.870 and 0.926, with slight fluctuations across
different batch sizes and optimizers. Similarly, the
F1 Score varies, demonstrating values from 0.635 to
0.879, emphasizing the trade-offs between precision
and recall.
When comparing Figures 4 and 5, the results show
that the (SMOTE + LSTM + FDL) model offers ex-
cellent overall performance, surpassing the results of
the (SMOTE+ CNN + FDL) model, in terms of recall
and F1 score. This makes it a recommended choice
for achieving robust and balanced performance.
Table 2: Performance Metrics for LSTM with Adam Opti-
mizer Across Different Batch Sizes.
Batch Size Accuracy (ACC) Precision Recall F1 Score
32 0.999 0.810 0.907 0.842
64 0.999 0.887 0.870 0.879
128 0.999 0.825 0.889 0.885
256 0.999 0.887 0.889 0.879
Table 3: Performance Metrics for CNN with Adam Opti-
mizer Across Different Batch Sizes.
Batch Size Accuracy Precision Recall F1 Score
32 0.999 0.830 0.889 0.844
64 0.998 0.470 0.926 0.610
128 0.998 0.500 0.870 0.635
256 0.999 0.880 0.815 0.846
The comparison between LSTM and CNN mod-
els with Adam optimizer and SMOTE technique re-
veals distinct performance characteristics across var-
ious batch sizes. The LSTM model consistently
achieves high accuracy, maintaining values at 0.999
across different batch sizes as observed in Table I . For
precision, LSTM ranges from 0.810 to 0.887, indicat-
ing reliable classification of positive cases. The recall
values for LSTM range from 0.870 to 0.907, show-
ing consistent ability to identify true positive cases.
Consequently, the F1 scores for LSTM vary between
0.842 and 0.885, reflecting a balanced performance
in terms of both precision and recall with Adam opti-
mizer.
In contrast, as displayed in Table II, the CNN
model shows variability in performance metrics
across batch sizes with Adam optimizer. While CNN
achieves high accuracy, ranging from 0.998 to 0.999,
its precision varies significantly, from 0.470 to 0.880.
This variability suggests different levels of effective-
ness in correctly identifying positive cases. CNN also
shows fluctuating recall values, ranging from 0.815
to 0.926, indicating varying success in capturing true
positive cases across different batch sizes. Corre-
spondingly, the F1 scores for CNN range widely from
0.610 to 0.846, highlighting its sensitivity to batch
size variations and the consequent impact on overall
model performance with Adam optimizer.
Overall, these results underscore the LSTM
model’s stability and balanced performance with
Adam optimizer and SMOTE, whereas the CNN
model’s effectiveness appears to be more sensitive to
batch size adjustments, potentially influencing its pre-
cision and recall outcomes significantly.
In addition, the comparison between LSTM and
CNN models using the SGD optimizer shows dis-
tinct performance profiles across batch sizes. As
seen in Table IV, LSTM achieves high accuracy
(0.998–0.999) with variable precision (0.520–0.646)
and stable recall (0.889–0.926), resulting in F1
Enhanced Credit Card Fraud Detection Using Federated Learning, LSTM Models, and the SMOTE Technique
373

Table 4: Performance Metrics for LSTM with SGD Opti-
mizer Across Different Batch Sizes.
Batch Size Accuracy Precision Recall F1 Score
32 0.998 0.530 0.926 0.736
64 0.999 0.610 0.907 0.826
128 0.999 0.520 0.889 0.644
256 0.999 0.646 0.926 0.762
Table 5: Performance Metrics for CNN with SGD Opti-
mizer Across Different Batch Sizes.
Batch Size Accuracy Precision Recall F1 Score
32 0.998 0.495 0.907 0.626
64 0.999 0.610 0.907 0.718
128 0.999 0.783 0.889 0.825
256 0.998 0.527 0.907 0.644
scores between 0.644 and 0.826. In contrast, CNN
achieves similar accuracy (0.998–0.999) but exhibits
more variability in precision (0.495–0.783) and recall
(0.889–0.907), with F1 scores ranging from 0.626 to
0.825, as shown in Table V. While LSTM shows con-
sistent performance, CNN’s variability highlights the
need for more tuning based on the application.
Table 6: Comparison between Previous Work and Our Best
System (Smote+LSTM+ADAM+FDL).
Model Accuracy Precision Recall F1 Score
GRU (Forough and Momtazi, 2022) 0.997 0.8626 0.7208 0.7792
LSTM (Forough and Momtazi, 2022) 0.997 0.8575 0.7408 0.7866
Ensemble model (Forough and Momtazi, 2022) 0.998 0.9569 0.6674 0.7813
Smote+CNN (H. Ghafoor and Khan, 2022) 0.998 0.8263 0.8095 0.8178
Smote+LSTM+ADAM+FDL 0.999 0.887 0.889 0.879
4.3 Comparison with Previous Work
As exhibited in Table VI, our proposed (SMOTE+
LSTM+ ADAM+ FDL) system achieves exceptional
accuracy (0.999), precision (0.887), recall (0.889),
and F1 Score (0.879), showcasing its effectiveness
in credit card fraud detection. In contrast, previ-
ous works such as the GRU (Forough and Mom-
tazi, 2022) and LSTM (Forough and Momtazi, 2022)
models reported precisions of 0.8626 and 0.8575 re-
spectively, with recall values of 0.7208 and 0.7408.
The ensemble model (Forough and Momtazi, 2022)
achieved a high precision of 0.9569 but lower recall
(0.6674) and comparable F1 Score (0.7813). More-
over, the SMOTE + CNN approach (H. Ghafoor and
Khan, 2022) demonstrated precision (0.8263), re-
call (0.8095), and F1 Score (0.8178), highlighting
different trade-offs in performance compared to our
method. These comparisons underscore the advance-
ments and efficacy of our secure and intelligent pro-
posed approach in addressing the challenges of credit
card fraud detection.
Figure 4: Performance Metrics vs Batch Size for
(SMOTE+LSTM+FDL) Model with Optimizers (ADAM
vs SGD).
Figure 5: Performance Metrics vs Batch Size for
(SMOTE+CNN+FDL) Model with Optimizers (ADAM vs
SGD).
5 CONCLUSION
This study compared LSTM and CNN models for
credit card fraud detection within a Federated Learn-
ing framework, using SMOTE to address class imbal-
ance. LSTM outperformed CNN in precision, recall,
and F1-score, achieving a precision of 0.887, recall
of 0.889, and F1-score of 0.879, compared to CNN’s
0.880, 0.815, and 0.846. These results highlight
LSTM’s effectiveness with sequential data. The com-
bination of LSTM, Federated Learning, Adam opti-
mizer, and SMOTE provided the best performance,
suggesting it as an optimal approach for fraud detec-
tion systems. Future work should focus on improving
Federated Learning protocols and exploring advanced
models.
REFERENCES
A. Singh, R. R. and Tiwari, A. (2021). Credit card fraud
detection under extreme imbalanced data: A compar-
ative study of data-level algorithms. J Exp Theor Artif
Intell, 34:1–28.
Akbar, M. and Chowanda, A. (2022). Sgd optimizer to re-
duce cost value in deep learning for customer churn
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
374

prediction. Journal of Theoretical and Applied Infor-
mation Technology, 100(11):8.
al., W. L. (2020). Federated learning in mobile edge net-
works: A comprehensive survey. IEEE Commun Surv
Tutor, 22(3).
al., X. L. (2017). Can decentralized algorithms outperform
centralized algorithms? a case study for decentralized
parallel stochastic gradient descent. In Adv Neural Inf
Process Syst, volume 30, page 6255.
al., Y. W. (2019). Ffd: A federated learning-based method
for credit card fraud detection. J Big Data, LNCS
11514:18–32.
Albertio, C. (2019). Towards efficient and privacy-
preserving federated deep learning. In International
Conference on Science and Technology on Communi-
cation Security Laboratory. IEEE.
Ali, A. H., Ammar, B., Charfeddine, M., and Hamed, B. B.
(2024a). Enhanced intrusion detection based hybrid
meta-heuristic feature selection.
Ali, A. H., Charfeddine, M., Ammar, B., and Hamed,
B. B. (2024b). Intrusion detection schemes based on
synthetic minority oversampling technique and ma-
chine learning models. In 2024 IEEE 27th Interna-
tional Symposium on Real-Time Distributed Comput-
ing (ISORC), pages 1–8.
Dornadula, V. and Geetha, S. (2019). Credit card fraud de-
tection using machine learning algorithms. In Proce-
dia Comput Sci, volume 165, pages 631–641.
Forough, J. and Momtazi, S. (2022). Ensemble of deep se-
quential models and machine learning techniques for
credit card fraud detection. Expert Systems with Ap-
plications, 197.
G. Bejjanki, G. J. and Narsimha, G. (2018). Class Process-
ing and Systems. Springer.
Group—ULB, M. L. (2018). Credit card fraud detection
anonymized credit card transactions labeled as fraud-
ulent or genuine.
H. Ghafoor, H. K. and Khan, M. (2022). Credit card fraud
detection with machine learning algorithms: A sys-
tematic literature review. Sustainability, 11.
Javatpoint (2024). Performance metrics in machine learn-
ing.
K. Chen, S. S. and Zhang, L. (2019). Big data–bigdata
2019: 8th international congress, held as part of the
services conference federation, scf 2019, san diego,
ca, usa, june 25–30, proceedings. volume 11514.
Springer.
Kamaruddin, S. and Ravi, V. (2016). Credit card fraud de-
tection using big data analytics: Use of psoaann-based
one-class classification. In Proceedings of the Interna-
tional Conference on Informatics and Analytics, pages
1–8, Pondicherry, India.
Li, X. and Zhao, L. (2021). Bank fraud detection using long
short-term memory networks with attention mecha-
nism. Expert Systems with Applications.
M. Fahmi, A. H. and Nagati, K. (2016). Data mining
techniques for credit card fraud detection: Empirical
study. In Sustain Vital Technol Eng Inf, pages 1–9.
Moneyzine (2023). Credit card fraud statistics & facts for
2023.
Payments, C. (2023). Credit card fraud statistics 2023.
Report, N. (2023). Global fraud losses 2023 projections.
ResearchGate (2022). The effect of imbalanced data on
credit card fraud detection.
Salam, M. A., Fouad, K. M., Elbably, D. L., and Elsayed,
S. M. (2023). Federated learning model for credit card
fraud detection with data balancing techniques. Neu-
ral Computing and Applications.
Stout, N. Undersampling and oversampling statistics visual
example. Pinterest.
Suvarna, R. and Kowshalya, A. M. (2020). Credit card fraud
detection using federated learning techniques. J Web
Eng Technol, 7(3):356–367.
V. Felbab, P. K. and Horv
´
ath, T. (2021). Optimization in
federated learning.
X. Niu, L. W. and Yang, X. (2019). A comparison study of
credit card fraud detection: Supervised versus unsu-
pervised. arXiv preprint arXiv:1904.10604.
X. Yao, T. Huang, C. W. R. Z. and Sun, L. (2019). Towards
faster and better federated learning: A feature fusion
approach. In 2019 IEEE International Conference on
Image Processing (ICIP), pages 175–195, Taipei, Tai-
wan. IEEE.
Zhang, W. and et al., T. W. (2021). Dynamic fusion-based
federated learning for covid-19 detection. In IEEE In-
ternet Things J, volume 8, pages 15884–15891.
Zhang, Y. and Liu, H. (2022). Convolutional neural net-
works for credit card fraud detection with imbalanced
datasets. IEEE Transactions on Neural Networks and
Learning Systems.
Enhanced Credit Card Fraud Detection Using Federated Learning, LSTM Models, and the SMOTE Technique
375