Synthetic Data Generation and Federated Learning as Innovative
Solutions for Data Privacy in Finance
Elif
¨
Ozcan
1,2 a
, Rus¸en Akkus¸ Halepmollası
1,2 b
and Yusuf Yaslan
1 c
1
Faculty of Computer and Informatics Engineering, Istanbul Technical University, Turkey
2
T
¨
UB
˙
ITAK Informatics and Information Security Research Center, Kocaeli, Turkey
Keywords:
Finance, Synthetic Data, Federated Learning, Artificial Intelligence.
Abstract:
Financial services generate vast, complex and diverse datasets, yet data privacy issues pose significant chal-
lenges for secure usage and collaborative analysis. Synthetic data generation can offer an innovative solution
while preserving privacy without exposing sensitive information. Also, federated learning enables collabo-
rative model training across clients while maintaining data privacy. In this study, we used Default Credit
Card dataset and employed diffusion based synthetic data generation to evaluate its impact on centralized and
federated learning approaches. To this end, we offer comprehensive benchmarking of synthetic, real, and
hybrid datasets by employing four machine learning classifiers both centrally and federated. Our findings
demonstrate that synthetic data effectively improves results, especially when combined with real data. We
also conduct client specific experiments in federated learning when addressing highly imbalanced or incom-
plete class distributions. Moreover, we evaluate FedF1 aggregation method, which aims to improve global
model performance by optimizing F1-score. To the best of our knowledge, this is the first study to integrate
synthetic data generation and federated learning on a financial dataset to provide valuable insights for secure
and collaborative learning.
1 INTRODUCTION
Artificial Intelligence (AI) has been a transformative
and innovative force in the financial sector, including
banking, insurance, trading, risk management, and
modern FinTech services (Cao, 2022). AI applica-
tions, particularly Machine Learning (ML) and Deep
Learning (DL), are crucial for modeling the complex
linear and nonlinear behaviors of financial variables
to address problems beyond the scope of traditional
models (Ahmed et al., 2022). Meanwhile, ML models
trained on vast amounts of financial data can achieve
higher scores in terms of evaluation metrics and en-
able more robust and efficient data driven decisions.
Financial services generate vast, complex and diverse
dataset; however, the sensitive and personally identi-
fiable features of financial data create significant chal-
lenges and limitations for its usage and sharing (As-
sefa et al., 2020).
One promising solution to handle data privacy and
security issues is synthetic data generation, which
a
https://orcid.org/0009-0002-3423-131X
b
https://orcid.org/0000-0002-9941-2712
c
https://orcid.org/0000-0001-8038-948X
mirrors the statistical properties and patterns of real
data without sharing sensitive data (Lu et al., 2023).
It aims to protect the privacy of customers due to laws
such as General Data Protection Regulation (GDPR)
(Hoofnagle et al., 2019) and Health Insurance Porta-
bility and Accountability Act (HIPAA) (Cohen and
Mello, 2018). Also, sharing realistic synthetic data
between institutions and within research commu-
nity allows training ML models on privacy-compliant
datasets and enables the development of effective so-
lutions to technical challenges. Moreover, synthetic
data can address the lack of historical data for cer-
tain events to provide counterfactual data for testing
strategies, as well as can handle class imbalances in
datasets to improve the performance of ML models,
particularly in cases like fraud detection (Assefa et al.,
2020).
Federated Learning (FL) emerges as another inno-
vative approach to address privacy concerns in finan-
cial data analysis by enabling multiple institutions to
collaboratively train ML models without the need to
share sensitive or raw data (Yang et al., 2019). Data
remains securely within the institutions’ premises,
and only model updates are shared. It ensures that
78
Özcan, E., Halepmollası, R. A. and Yaslan, Y.
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance.
DOI: 10.5220/0013440900003956
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 7th International Conference on Finance, Economics, Management and IT Business (FEMIB 2025), pages 78-89
ISBN: 978-989-758-748-1; ISSN: 2184-5891
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

privacy is maintained, as the underlying financial data
never leaves the organization, thus complying with
regulatory constraints such as GDPR (Truong et al.,
2021). FL also facilitates data collaboration across in-
stitutions and allows them to leverage diverse datasets
to improve model results without violating privacy
policies (Mothukuri et al., 2021).
Considering the aforementioned data privacy and
security issues, in this study, we employed two inno-
vative approaches: (i)Synthetic data generation and
(ii)FL. We leveraged a diffusion model for synthetic
data generation and explored its impact on both cen-
tralized and FL approaches across several ML al-
gorithms, including Logistic Regression (LR), Sup-
port Vector Classifier (SVC), Stochastic Gradient De-
scent Classifier (SGDC), and Multi Layer Percep-
tron (MLP). We benchmark and evaluate the real,
synthetic and hybrid data in centralized and FL ap-
proaches under six distinct experimental scenarios.
Additionally, we conducted two case studies to an-
alyze specific challenges. Case Study 1 focused on
the impact of FL and synthetic data at client level
and Case Study 2 focused on addressing highly im-
balanced and incomplete class distributions. In this
context, our contributions are as follows:
• We present comprehensive benchmarking of syn-
thetic, real and hybrid data in both centralized
and FL environments. To the best of our knowl-
edge, this is the first study to comprehensively in-
vestigate the integration of synthetic data genera-
tion and FL approach using Default Credit Card
dataset to address critical issues in data privacy,
accessibility, and class imbalance.
• We introduce a client-level analysis in FL to in-
vestigate whether it improves model outputs, par-
ticularly in scenarios with imbalanced or incom-
plete class distributions.
• We evaluate a novel FedF1 aggregation method
(Aktas¸ et al., 2024) to optimize global model per-
formance in FL to explore its ability to handle het-
erogeneity and imbalance clients.
Our contributions provide a robust framework for
integrating synthetic data generation and FL approach
in financial applications to address data privacy, secu-
rity and accessibility issues.
Structure of the Paper. Section 2 summarizes previ-
ous works on synthetic data generation and FL ap-
proaches in finance. In Section 3, we present the
methodology. Section 4 describes the dataset and ex-
perimental setup of case studies. In Section 5, the re-
sults of the study are reported and discussed. Section
6 concludes the paper and offers future work.
2 LITERATURE REVIEW
In this section, we review existing literature on syn-
thetic data generation and FL in the financial domain.
This review is organized into two subsections: the
role of synthetic data in financial applications and ad-
vancements in federated learning for finance.
2.1 Synthetic Data Generation in
Finance
Synthetic data generation plays a crucial role in fi-
nancial applications by addressing various issues such
as data scarcity, class imbalance, and privacy con-
straints. Khaled et al. (Khaled et al., 2024) explored
the use of synthetic data to improve ML models for
credit card fraud detection. Authors employed the
SMOTE technique to address the severe class imbal-
ance in financial datasets, where fraudulent transac-
tions are significantly underrepresented. By train-
ing ML models on the generated synthetic data, they
observed notable improvements in accuracy and re-
call, particularly in detecting minority class detection.
This research underscores the potential of synthetic
data to mitigate data imbalance challenges and im-
prove the performance of fraud detection models in
the financial sector.
Building on the promise of synthetic data,
Jolicoeur-Martineau et al. (Jolicoeur-Martineau et al.,
2023) proposed a novel framework that integrates
score-based diffusion models with conditional flow
matching for tabular data generation and imputation
by using XGBoost. That approach is specifically
designed to handle mixed-type tabular data, includ-
ing both categorical and numerical features, a com-
mon challenge in tabular data modeling. Through ex-
tensive experimentation on 27 datasets from diverse
domains, the method demonstrated superior perfor-
mance in data generation tasks compared to state-of-
the-art DL-based generative models while maintain-
ing competitive results in data imputation scenarios.
Additionally, key advantage of the proposed approach
is its efficiency, as it can leverage parallel CPU train-
ing and bypass the need for computationally expen-
sive GPUs. This work highlights the potential of com-
bining advanced generative modeling techniques with
traditional ML algorithms to effectively address tabu-
lar data challenges.
Furthermore, Sattarov et al. (Sattarov et al., 2023)
introduced FinDiff, a novel diffusion-based model
specifically designed to generate synthetic tabular
data in the financial domain. The model addresses
the challenges associated with mixed-type data, such
as the coexistence of numerical and categorical fea-
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
79

tures. FinDiff was rigorously evaluated on three real-
world financial datasets and focused on regulatory
tasks including economic scenario modeling, stress
testing, and fraud detection—key applications in fi-
nance where data availability and privacy are critical
concerns. Their experimental results demostrated that
FinDiff can preserve the statistical properties of the
original dataset and generate high-fidelity synthetic
data while ensuring utility and privacy. Thus, the
model offers a robust solution to data-sharing chal-
lenges in the financial industry. Moreover, authors
highlighted the versatility of FinDiff in supporting
downstream ML tasks and showing competitive per-
formance compared to traditional methods. FinDiff
not only enhances data accessibility but also aligns
with the regulatory requirements of the financial sec-
tor. Therefore, it can be a valuable tool for FL appli-
cations.
2.2 Federated Learning in Finance
FL has received significant attention from researchers
and practitioners as it enables collaborative model
training without sharing sensitive data (
¨
Ulver et al.,
2023)(Zhang et al., 2021)(Yurto
˘
glu et al., 2024).
Wang et al. (Wang et al., 2024) proposed Feder-
ated Knowledge Transfer (FedKT), a FL approach de-
veloped for credit scoring while preserving data pri-
vacy. The approach enables collaboration among fi-
nancial institutions without sharing raw data to ad-
dress privacy concerns in credit scoring. A key chal-
lenge in FL is the heterogeneity of data distribu-
tions across participants, which can hinder the learn-
ing capacity of the global model. For this purpose,
FedKT combines fine-tuning and knowledge distilla-
tion techniques to effectively extract general knowl-
edge from the global model’s early layers and task-
specific knowledge from its outputs. Experimental
evaluations on four distinct credit datasets demon-
strated that FedKT outperforms existing FL algo-
rithms in terms of predictive performance and robust-
ness. Its ability to balance privacy preservation with
high model performance makes it particularly valu-
able in the financial sector, where data sensitivity and
regulatory compliance are critical.
In addition to privacy concerns, data imbalance
poses a significant challenge in FL environments.
Zhang et al. (Zhang et al., 2024) explored the chal-
lenges posed by data imbalance in FL for credit
risk forecasting, a critical task in financial decision-
making. They analyzed the performance of three ML
models—Multilayer Perceptron (MLP), Long Short-
Term Memory (LSTM), and eXtreme Gradient Boost-
ing (XGBoost)—across multiple datasets with vary-
ing client numbers and data distribution patterns.
They achieved an average performance improvement
of 17.92% and their findings revealed that FL mod-
els significantly outperformed local models for non-
dominant clients with smaller, highly imbalanced
datasets. However, for dominant clients with larger
datasets, FL models offered no clear advantage over
local models, thus, authors highlighted potential dis-
incentives for their participation. The study empha-
sized the need for strategies to mitigate the effects of
data imbalance and ensure equitable benefits for all
participants in FL environments.
Trust and interpretability are also critical for
the adoption of FL in finance. Awosika et al.
(Awosika et al., 2023) introduced a novel approach
that combines FL and eXplainable Artificial Intelli-
gence (XAI) to enhance financial fraud detection sys-
tems. FL enables multiple financial institutions to
collaboratively train a shared fraud detection model
without exchanging sensitive customer data, thereby
upholding data privacy and confidentiality. The in-
tegration of XAI ensures that the model’s predic-
tions are interpretable by human experts and fosters
transparency and trust in the system. Authors con-
ducted experiments on realistic transaction datasets
and demonstrated that the FL-based fraud detection
system consistently achieved high performance met-
rics. They underscored FL’s potential as an effec-
tive and privacy-preserving tool in combating finan-
cial fraud.
3 METHODOLOGY
In this section, we provide our methodology that in-
volves synthetic data generation, FL approach and
ML algorithms.
3.1 Synthetic Data Generation
In this study, we follow synthetic data production pro-
cedure presented in FinDiff: Diffusion Models for
Financial Tabular Data Production (Sattarov et al.,
2023). For this purpose, we used Gaussian Diffu-
sion Models to generate synthetic data customized to
mixed-type tabular datasets, which are common in fi-
nancial applications. The methodology is intended to
overcome the issues of working with heterogeneous
data that contains both numerical and category vari-
ables.
Gaussian Diffusion Models operate by gradually
transforming data distributions through a two-step
process. In the forward diffusion phase, Gaussian
noise is incrementally added to the original data,
FEMIB 2025 - 7th International Conference on Finance, Economics, Management and IT Business
80

effectively smoothing its complex structure into a
noise-dominated state. Thus, the learning of high-
dimensional relationships within the data is facili-
tated. In the reverse diffusion phase, noise is sys-
tematically removed and reconstructed synthetic sam-
ples that approximate the original data distribution. A
learned score function guides the reverse process to
ensure that the generated data aligns closely with the
original dataset’s statistical and structural properties
(Ho et al., 2020)(Sohl-Dickstein et al., 2015).
In line with the methodology presented in (Sat-
tarov et al., 2023), we also evaluated the quality
and utility of the generated synthetic data using sev-
eral metrics, including fidelity, utility, synthesis, and
privacy. Fidelity measures how well the synthetic
data replicated the statistical properties of the orig-
inal data, both at the column level (individual fea-
tures) and row level (holistic data structures). Util-
ity evaluates the ability of synthetic data to support
downstream ML tasks such as fraud detection and
credit scoring. Synthesis ensures that the generated
data maintained structural alignment with the original
dataset. Privacy assesses resistance to privacy attacks
such as membership inference.
3.2 Federated Learning
FL is a decentralized ML approach designed for train-
ing models collaboratively across multiple clients
while preserving data privacy. Unlike traditional cen-
tralized approaches, where data is collected and pro-
cessed on a central server, FL ensures that data re-
mains on the clients. Only model updates, such
as gradients and weights, are shared with a central
server, in which the model parameters are aggregated
to create a global model. As shown in the Figure 1,
each client trains the model locally on its own dataset
to create a global model without sharing the data and
ensure that sensitive data never leaves the clients’ en-
vironment. After local training, each client sends only
the model parameters to the server that aggregates the
weights from all clients to update the global model.
The updated global model is shared with all clients,
and the process is repeated for several iterations until
the model converges.
In this study, we employed two aggregation meth-
ods. The first method, namely FedAvg, computes a
weighted average of model updates from participat-
ing clients based on their dataset sizes to ensure pro-
portional contribution to the global model (McMahan
et al., 2017). The second method, namely FedF1, ag-
gregates model updates by assigning weights based
on the clients’ F1-scores (Aktas¸ et al., 2024). FedF1
prioritizes contributions from clients with higher F1-
scores to reflect more stable and accurate local mod-
els. In other words, by using F1-scores as a weighting
factor, it aims to improve the overall performance and
reliability of the global model, particularly in scenar-
ios with imbalanced datasets or heterogeneous client
performance.
Figure 1: Synthetic Data Generation.
3.3 Machine Learning Algorithms
When comparing the both synthetic−real data and
central−decentral approaches, we examined the fea-
sibility of using four classifiers, namely LR, SVC,
SGDC and MLP. The employed classifiers are, in
short, described below:
Logistic Regression. (Hosmer et al., 2013) models
the probability of a binary class label using the sig-
moid function by transforming a linear combination
of input features into a probability score. It is a para-
metric and discriminative method and focuses on the
direct mapping between features (independent vari-
ables) and class labels (dependent variables). In this
study, we conducted hyperparameter tuning using the
training and validation datasets, resulting in the best
parameters: C = 0.01, max iter = 5000, penalty = l1,
solver = saga, and class weight = balanced.
Support Vector Classifier. (Cortes and Vapnik,
1995) constructs a hyperplane that separates classes
in the feature space with maximum margin. In this
study, we employed a linear kernel to model linearly
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
81

separable data. Hyperparameter tuning focused on
improving the model’s handling of class imbalances
and convergence properties, with the optimal param-
eters identified as C = 0.1, max iter = 1000, and
class weight = balanced.
Stochastic Gradient Descent Classifier. (Bottou,
2010) is a linear classifier that leverages Stochastic
Gradient Descent for optimization. It iteratively up-
dates the model parameters by computing the gra-
dient of the loss function with respect to a single
training example to make it highly efficient for large-
scale and sparse datasets. In this study, we fine-
tuned the hyperparameters, and determined the op-
timal configuration as al pha = 0.1 (regularization
term), max iter = 5000 (maximum number of itera-
tions), penalty = elasticnet (combination of L1 and
L2 regularization), and l1 ratio = 0.5 (balance be-
tween L1 and L2 regularization).
Multi-Layer Perceptron. (Goodfellow et al.,
2016) is a feedforward neural network that cap-
tures non-linear relationships between input fea-
tures and target labels using multiple layers of
neurons. Training is performed using backprop-
agation to optimize the weights of the network.
The hyperparameter tuning process determined
the best configuration as activation = relu,
al pha = 0.001, hidden layer sizes = (50, ),
learning rate = adaptive, max iter = 200, and
solver = adam.
4 EXPERIMENTAL SETUP
In this section, we present the details of the dataset we
utilized and the experiments we conducted, including
the training configuration, evaluation scenarios, case
studies, and evaluation metrics. The overall flow of
the experimental setup is shown in Figure 2.
4.1 Dataset
In this study, we used Default of Credit Card
Clients(DCCC) dataset (Yeh and Lien, 2009), ob-
tained from the UCI Machine Learning Repository.
DCCC dataset includes 30,000 records of credit card
clients in Taiwan and includes both categorical and
numerical features. It provides a comprehensive set of
attributes, including demographic information, pay-
ment history, bill statements, and a default payment
indicator. It is complete, with no missing values, and
the target variable is binary. The dataset has a class
imbalance, with class ratio of 3:1.
4.2 Training Configuration
We randomly split 90% of the data for training and
10% for testing. Random splitting of the dataset can
lead to significant variations in the target variable ra-
tios, which may impact model performance, espe-
cially since our dataset is imbalanced with a small
number of samples in the default class. To address
this issue, we ensured that the data split was per-
formed with stratification. To obtain more reliable
results, we repeated the experiments 10 times, each
with a different random seed to shuffle the order of
the samples, and calculated the average performance
scores across all runs.
We utilized the training data to generate synthetic
data from the training set. For this purpose, we em-
ployed the diffusion model, as detailed in Section 3.
Moreover, we partitioned the training data equally
into five subsets, representing five distinct clients for
FL setup. We built federated models using the real
train set distributed across those clients and evaluated
models’ performance using the real test set. When
exploring the potential of synthetic data in FL, we in-
dependently applied the same diffusion model to each
client to generate client-specific synthetic data. Please
note that we generated synthetic data for each client
using only their local data, as clients in real-world FL
scenarios cannot access to each other’s data. Over-
all, to evaluate the impact of data augmentation on
model performance, we trained federated models on
three data types: real data, synthetic data and hybrid
data which is a combination of real and synthetic data.
Furthermore, to compare the FL approach with the
centralized approach, we used the same train-test con-
figurations when building ML models centrally. Sim-
ilar to the FL approach, we trained centralized models
on three data types: real data, synthetic data and hy-
brid data.
Table 1 provides a detailed summary of the data
distribution across centralized and federated setups to
highlight the class distribution in each subset. In all
scenarios, we used only the test set split from the real
data for evaluation. In other words, model testing
across all configurations was performed on the real
test set. Thus, we ensured consistency and compara-
bility across centralized and federated setups, as well
as across models trained on real, synthetic, and com-
bined datasets.
FEMIB 2025 - 7th International Conference on Finance, Economics, Management and IT Business
82

Figure 2: Illustration of proposed experimental setup.
Table 1: Data distribution across clients and test set.
# Samples
Class 0
# Samples
Class 1
Class 1
Proportion (%)
Central Data 21023 5977 22.1
Client 1 4229 1171 21.6
Client 2 4209 1191 22.0
Client 3 4218 1182 21.8
Client 4 4187 1213 22.4
Client 5 4180 1220 22.5
Test Data 2341 659 21.9
4.3 Benchmarking: Evaluating
Synthetic Data in Centralized and
FL Approaches
To investigate the impact of synthetic data on FL, we
conducted a comprehensive analysis using four al-
gorithms: Logistic Regression (LR), Support Vector
Classifier (SVC), Stochastic Gradient Descent Clas-
sifier (SGDClassifier) and Multi-Layer Perceptron
(MLP). Our goal is to evaluate the performance of
both centralized and federated models on varying data
types and provide insights into how synthetic data in-
fluences learning outcomes. Also, we designed the
experiments around six distinct scenarios:
• Central+Real Data. In this scenario, centralized
models are trained solely on the real data. The
objective is to assess the baseline performance of
centralized models without the influence of syn-
thetic data.
• Central+Synthetic Data. In this scenario, central-
ized models are trained using only synthetic data
generated from the training set. This allows us to
evaluate the impact of synthetic data in a central-
ized learning environment by comparing perfor-
mance with the real data scenario.
• Central+Hybrid Data. This scenario involves
centralized models trained on a combination of
real and synthetic data. The goal is to assess the
effectiveness of data augmentation.
• FL+Real Data. In this scenario, federated models
are trained using only real data distributed across
clients. This scenario provides a baseline for FL
performance with real data.
• FL+Synthetic Data. In this scenario, federated
models are trained using synthetic data generated
for each client. The purpose is to explore the
potential of synthetic data in a FL environment
and evaluate how it influences model performance
compared to real data.
• FL+Hybrid Data. This scenario involves feder-
ated models trained on a combination of real and
synthetic data. By incorporating both types of
data, the setup evaluates the impact of data aug-
mentation on FL performance, similar to the hy-
brid data scenario in centralized models.
4.4 Case Study 1: Evaluating FL
Performance with Synthetic Data
To further explore the benefits of FL at the client level,
we conducted an additional case study focusing on
three selected clients (Client 1, Client 3, and Client 5).
We evaluated their performance under various config-
urations to gain a deeper understanding of the impact
of FL and the use of synthetic data at the client level.
We utilized SVC and MLP based on the benchmark-
ing results of which MLP achieved the highest accu-
racy scores and SVC reached the best F1-scores.
For each client, we performed experiments us-
ing locally trained centralized models on three dif-
ferent data types: real data, synthetic data, and hy-
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
83

brid data. We trained the models on the individual
client’s data to represent a baseline for local training
without federated collaboration. In FL setup, we con-
ducted each model training over 10 rounds using the
FedAvg aggregation method. We evaluated two dis-
tinct model types at the end of the training process to
assess the impact of global collaboration and client-
specific fine-tuning:
• Global Federated Model: This model represents
the aggregated global model produced by the
server after the 10th communication round. It re-
flects the combined knowledge learned from all
participating clients.
• Client-Adapted Federated Model: This model is
derived from the global federated model after the
10th round but is further fine-tuned locally on
each client’s own data.
With the above model types, we aim to explore,
in detail, the performance improvements that FL can
bring to local clients when data centralization is not
feasible due to privacy concerns, regulatory restric-
tions, or operational constraints.
4.5 Case Study 2: Addressing
Imbalanced Class Distribution
In this case study, we investigated the potential of FL
to address a scenario where clients have highly imbal-
anced or incomplete class distributions. To simulate
such a scenario, we created three clients: one client
with no samples labeled as class 1 and two clients
with balanced class distributions. The setup reflects
real-world situations, such as a bank branch with no
recorded fraud cases, while other branches have suf-
ficient data for both classes. The data distribution
among the clients and the test set is summarized in
Table 2.
Table 2: Data distribution across clients and test set for Case
Study 2.
Class 0 Class 1
Client 1 5976 0
Client 2 2988 2988
Client 3 2988 2988
Test Data 659 659
When the first client trains a local model indepen-
dently, it has no knowledge of class 1 due to the ab-
sence of positive samples on its dataset. Therefore, its
local model can be incapable of predicting the posi-
tive class. On the other hand, by participating in FL,
that client can leverage knowledge aggregated from
other clients and gain access to information about
class 1 without sharing its raw data, thus preserving
data privacy.
Additionally, we explored the limitations of syn-
thetic data generation in this context. Even if syn-
thetic data were generated for the first client, the ab-
sence of positive samples would prevent the genera-
tion of meaningful data for class 1. Thus, it highlights
a critical scenario where FL provides a unique advan-
tage over local models and synthetic data augmenta-
tion.
For aggregation in this study, we employed the
FedF1 method, designed to optimize the federated
model for imbalanced data scenarios. By focusing
on F1-score optimization during aggregation, FedF1
ensures that the global model performs effectively
across clients with differing class distributions.
5 ANALYSIS OF THE RESULTS
In this study, we evaluate the performance of synthetic
data in centralized and FL approaches using four ML
models, including LR, SVC, SGDC, and MLP. In
Case Study 1, we investigate FL performance with
synthetic data. With Case Study 2, we address im-
balanced class distribution using FL approach. We
evaluated the synthetic data generated for centralized
training to assess its similarity to the real data. We
also compared feature distributions and inter-feature
relationships between the real and synthetic datasets.
Figure 3 shows the probability distributions of se-
lected features (e.g., Gender, Pay0, Age, and Lim-
itbal) and illustrate a close match between syn-
thetic dataset and real dataset. Furthermore, Figure
4 presents the column pair trends and correlations
and also demonstrate strong consistency across both
datasets. The evaluation using the SDV framework
provided additional metrics to quantify the quality of
the synthetic data as follows:
• Column Shapes Score: 92.27%
• Column Pair Trends Score: 77.29%
• Overall Quality Score: 84.78%
In the subsections, we discuss, in detail, the ef-
fect of the different data types, FL and centralized
approaches, selected ML algorithms, and imbalanced
class distribution.
FEMIB 2025 - 7th International Conference on Finance, Economics, Management and IT Business
84

Figure 3: Probability distributions of selected features for real and synthetic data.
Table 3: Comparative analysis of different algorithms for Central and FL models.
SGD SVC LR MLP
F1 Acc Recall Prec F1 Acc Recall Prec F1 Acc Recall Prec F1 Acc Recall Prec
Central - Real Data 0.4926 0.6320 0.55 0.56 0.5977 0.6540 0.64 0.60 0.5699 0.6913 0.57 0.56 0.4922 0.7713 0.52 0.57
Central - Synthetic Data 0.4852 0.5804 0.55 0.55 0.6386 0.7190 0.66 0.63 0.5797 0.7003 0.58 0.57 0.4548 0.7723 0.50 0.53
Central - Hybrid Data 0.4939 0.6080 0.55 0.58 0.6335 0.7057 0.66 0.62 0.5772 0.6960 0.58 0.57 0.4667 0.7795 0.51 0.62
FL - Real Data 0.4573 0.5589 0.53 0.54 0.6037 0.6637 0.65 0.60 0.5695 0.6943 0.57 0.56 0.3944 0.5684 0.48 0.49
FL - Synthetic Data 0.3829 0.4561 0.52 0.48 0.6674 0.7980 0.65 0.70 0.5532 0.6913 0.55 0.55 0.3954 0.5055 0.49 0.49
FL - Hybrid Data 0.4009 0.4601 0.53 0.54 0.6447 0.7370 0.66 0.63 0.5568 0.6960 0.56 0.55 0.3042 0.3446 0.48 0.51
Figure 4: Column pair trends for real and synthetic data.
5.1 Benchmarking: Evaluating
Synthetic Data in Centralized and
FL Approaches
In this subsection, we compare the performance of
centralized and FL approaches using real data, syn-
thetic data and hybrid data.
The results presented highlight an insight into the
effectiveness of synthetic data in ML models, both in
centralized and FL frameworks. One of the primary
goals of this study was to evaluate whether models
trained on synthetic data could achieve performance
comparable to, or even exceed, those trained on real
data. From Table 3, we observe that while different
algorithms show varying levels of performance, the
use of synthetic data produces results that are com-
parable to—and occasionally better than—those ob-
tained using real data. Figure 5 visually illustrates
this observation, showcasing the performance distri-
bution of centralized and FL models across different
data configurations. The violin plots clearly demon-
strate that synthetic data yields performance distribu-
tions comparable to those of real data, further high-
lighting its effectiveness in both centralized and FL
settings.For instance, in the centralized models, the
SVC algorithm achieved an F1-score of 0.6386 with
synthetic data, outperforming the corresponding F1-
score of 0.5977 when trained on real data. A similar
trend is observed in FL models, where the synthetic
data-based SVC achieved an F1-score of 0.6674, ex-
ceeding the F1-score of 0.6037 obtained with real
data.
To delve deeper into the comparative performance
of the datasets, Table 4 provides a focused analysis of
F1-scores across all experimental setups. This table
shows that in both centralized and FL models, syn-
thetic data consistently performs competitively. Al-
though hybrid datasets often improve model perfor-
mance compared to real data (e.g., the F1-score of
0.4939 for the SGD method in centralized settings or
the F1-score of 0.6447 for the SVC algorithm in FL),
the gains over synthetic data are typically insignifi-
cant. This indicates that using synthetic data alone is
usually enough to produce competitive results, even
though hybrid data may be useful in certain situa-
tions. In many instances, hybrid data perform simi-
larly to synthetic data, suggesting that the additional
complexity of integrating real data may not always be
necessary.
The results underline that the quality of model pre-
dictions is unaffected by the use of synthetic data.
In contrast, synthetic data’s competitive performance
shows that it can potentially ease privacy issues, since
using the synthetic data provides similar results with-
out compromising the privacy of sensitive data in both
centralized and FL environments. The success of syn-
thetic data in achieving comparable or superior per-
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
85

Figure 5: Performance distribution of centralized and federated models across different data configurations.
Table 4: F1-Scores Across Centralized and Federated Learning Models Using Real, Synthetic, and Hybrid Data
Centralized Models FL Models
Real Data Synthetic Data Hybrid Data Real Data Synthetic Data Hybrid Data
SGD 0.4926 0.4852 0.4939 0.4573 0.3829 0.4009
SVC 0.5977 0.6386 0.6335 0.6037 0.6674 0.6447
LR 0.5699 0.5797 0.5772 0.5695 0.5532 0.5568
MLP 0.4922 0.4548 0.4667 0.3944 0.3954 0.3042
formance reaffirms its value as an alternative, espe-
cially in domains where real data is limited, sensitive,
or inaccessible.
5.2 Case Study 1: Evaluating FL
Performance with Synthetic Data
In this subsection, we compare the performance of
centralized and FL approaches across different clients
for FL global models, as well as fine-tuned FL mod-
els.
The results presented in Table 5 and Table 6 pro-
vide insights into the client-level impact of FL models
compared to centralized models. From the results, it
is evident that FL models generally achieve perfor-
mance that is comparable to, or even exceeds, that of
centralized models across most clients. For instance,
in Table 5, the FL global model for SVC achieved
an F1-score of 0.6649 for Client 2, outperforming the
centralized model’s F1-score of 0.5603. Similarly, in
Table 6, for the MLP model, the FL model achieved
an F1-score of 0.4673 for Client 2, which is compara-
ble to the centralized model’s F1-score of 0.4673. The
results show the effectiveness of FL in maintaining or
improving performance at the client level, ensuring
that models trained in a decentralized manner are ca-
pable of matching the results of centralized training.
5.3 Case Study 2: Addressing
Imbalanced Class Distributions
In this case study we focused on the limitations of
synthetic data in addressing class imbalance, partic-
ularly when certain classes are entirely absent from
a client’s local dataset. In such scenarios, synthetic
data generation alone fails to resolve the issue, as it
relies solely on the local data distribution and cannot
create representations for missing classes. FL, how-
ever, overcomes this limitation by aggregating knowl-
edge from multiple clients, enabling the global model
to learn from distributed datasets where the missing
class is present.
Table 7 highlights the performance of centralized
models and FL strategies, FedAvg and FedF1. The FL
FEMIB 2025 - 7th International Conference on Finance, Economics, Management and IT Business
86

Table 5: Client-Specific Evaluation of SVC in Centralized and Federated Learning Models.
Client 1 Client 2 Client 3
Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy
Central - Real Data 0.6016 0.6428 0.5939 0.6510 0.5868 0.6261 0.5603 0.6020 0.6263 0.6723 0.6275 0.6893
FL - Real Data 0.6016 0.6428 0.5939 0.6510 0.5868 0.6261 0.5603 0.6020 0.6263 0.6723 0.6275 0.6893
FL Global - Real Data 0.6078 0.6498 0.6037 0.6637 0.6078 0.6498 0.6037 0.6637 0.6078 0.6499 0.6037 0.6637
Central - Synthetic Data 0.6792 0.6431 0.6558 0.7867 0.5948 0.6188 0.5988 0.6867 0.7117 0.6099 0.6270 0.8003
FL - Synthetic Data 0.6792 0.6436 0.6576 0.7867 0.5948 0.6188 0.5988 0.6867 0.7117 0.6099 0.6270 0.8003
FL Global - Synthetic Data 0.6996 0.6514 0.6673 0.7980 0.6996 0.6514 0.6673 0.7980 0.6996 0.6514 0.6673 0.7980
Central - Hybrid Data 0.6215 0.6546 0.6277 0.7060 0.5761 0.6092 0.5579 0.6097 0.6895 0.6594 0.6710 0.7917
FL - Hybrid Data 0.6215 0.6546 0.6277 0.7060 0.5761 0.6092 0.5579 0.6097 0.6895 0.6594 0.6709 0.7917
FL Global - Hybrid Data 0.6373 0.6581 0.6447 0.7370 0.6373 0.6581 0.6447 0.7370 0.6373 0.6581 0.6447 0.7370
Table 6: Client-Specific Evaluation of MLP in Centralized and Federated Learning Models.
Client 1 Client 2 Client 3
Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy
Central - Real Data 0.5190 0.5035 0.4679 0.7645 0.4982 0.5036 0.4673 0.7574 0.5240 0.5107 0.4850 0.7549
FL - Real Data
0.5406 0.5286 0.5248 0.7030 0.5557 0.5378 0.5286 0.7299 0.5588 0.5617 0.5380 0.6671
FL Global - Real Data 0.4922 0.4833 0.3944 0.5684 0.4922 0.4833 0.3944 0.5684 0.4922 0.4833 0.3944 0.5684
Central - Synthetic Data 0.4546 0.4974 0.4461 0.7687 0.4894 0.5014 0.4595 0.7644 0.4510 0.4903 0.4478 0.7524
FL - Synthetic Data 0.4941 0.4988 0.4688 0.7433 0.5094 0.5130 0.4904 0.7067 0.4985 0.5031 0.4770 0.7266
FL Global - Synthetic Data 0.4871 0.4920 0.3954 0.5055 0.4871 0.4920 0.3954 0.5055 0.4871 0.4920 0.3954 0.5055
Central - Hybrid Data 0.5414 0.5044 0.4631 0.7689 0.4997 0.5035 0.4634 0.7643 0.4828 0.4972 0.4632 0.7515
FL - Hybrid Data 0.5607 0.5496 0.5271 0.6910 0.5415 0.5265 0.4974 0.6906 0.5580 0.5752 0.5482 0.6348
FL Global - Hybrid Data
0.5093 0.4869 0.3042 0.3446 0.5093 0.4869 0.3042 0.3446 0.5093 0.4869 0.3042 0.3446
Table 7: Case 2 - Comparison of Centralized Learning and Federated Learning with Different Strategies.
F1 Precision Recall Accuracy
Real Synthetic Hybrid Real Synthetic Hybrid Real Synthetic Hybrid Real Synthetic Hybrid
Central 0.3333 0.3333 0.3333 0.7500 0.7500 0.7500 0.5000 0.5000 0.5000 0.5000 0.5000 0.5000
FL Avg 0.3333 0.3333 0.3333 0.7500 0.7500 0.7500 0.5000 0.5000 0.5000 0.5000 0.5000 0.5000
FL Avg Global 0.3891 0.3894 0.3544 0.5399 0.5656 0.5913 0.5013 0.5015 0.5035 0.50013 0.5001 0.5035
FL FedF1 0.3333 0.3333 0.3333 0.7500 0.7500 0.7500 0.5000 0.5000 0.5000 0.5000 0.5000 0.5000
FL Global FedF1 0.4000 0.4003 0.4118 0.4952 0.4894 0.5998 0.4996 0.4799 0.4937 0.4996 0.4799 0.4937
Global FedF1 model achieves an F1-score of 0.4000
for real data, outperforming the centralized model
(F1-score: 0.3333). This demonstrates FL’s ability to
leverage data from other clients to predict underrep-
resented classes effectively, a capability that synthetic
data alone cannot provide.
Moreover, the FedF1 strategy proves more effec-
tive than FedAvg by prioritizing F1-scores during ag-
gregation, thereby enhancing the global model’s abil-
ity to handle imbalanced data. For example, when
using hybrid data (real + synthetic), the FL Global
FedF1 model achieves a F1-score of 0.4118, com-
pared to 0.3544 for FedAvg. These results underscore
the critical role of FL in scenarios where local data
distributions are severely imbalanced.
These results demonstrate that FL, particularly
with the FedF1 strategy, effectively addresses class
imbalance in decentralized environments. Unlike
synthetic data generation, which is constrained by
local data distributions, FL aggregates distributed
knowledge across clients, enabling robust model
training even in the absence of certain classes. This
highlights FL’s potential as a practical approach for
scenarios where class imbalance cannot be resolved
through conventional means.
6 FINAL REMARKS
This study highlights the potential of combining syn-
thetic data and FL to address critical challenges in
ML, such as data privacy, class imbalance, and decen-
tralized learning. Synthetic data emerged as a reliable
alternative to real data, consistently achieving com-
parable performance across both centralized and FL
settings. Its effectiveness underscores its applicabil-
ity for applications where privacy or data accessibil-
ity constraints make the use of real data impractical.
These results are particularly relevant for applications
in financial institutions, where data privacy regula-
tions prevent direct data sharing between entities, the
combination of synthetic data and FL enables collab-
orative learning without compromising sensitive in-
formation, ensuring that institutions can achieve per-
formance comparable to centralized models without
exposing their data.
At the same time, FL demonstrated its capacity to
enhance model robustness by leveraging distributed
knowledge across clients. This capability proved par-
ticularly crucial in scenarios where synthetic data
alone was insufficient, such as when certain classes
were entirely absent from a client’s local dataset. By
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
87

integrating data from other clients, FL effectively mit-
igated these limitations, enabling the global model to
address class imbalances and improve prediction ac-
curacy. Compared to traditional centralized learning
approaches, this combination not only preserves data
privacy but also enhances model robustness by lever-
aging distributed knowledge, making it particularly
effective in scenarios with class imbalances or miss-
ing labels.
The findings suggest that synthetic data and FL
are not only complementary but also mutually rein-
forcing. Synthetic data provides the foundation for
privacy-preserving ML, while FL extends this foun-
dation to handle more complex challenges inherent
in decentralized environments. Together, these ap-
proaches form a robust framework for developing
high-performing and privacy-conscious ML models
suitable for real-world applications.
For future work, there are several potential direc-
tions to build upon our current findings. First, the im-
pact of alternative synthetic data generation methods
could be examined, focusing on how different tech-
niques influence model performance in both central-
ized and FL frameworks. Furthermore, expanding the
scope of the study to include diverse datasets from
various domains would help validate the robustness
and applicability of the proposed approach. Another
promising avenue involves testing more advanced
classification algorithms to explore their potential for
improving both predictive accuracy and generaliza-
tion across heterogeneous environments. These direc-
tions would collectively contribute to a deeper under-
standing of the interplay between synthetic data and
FL in addressing real-world ML challenges.
REFERENCES
Ahmed, S., Alshater, M. M., El Ammari, A., and Ham-
mami, H. (2022). Artificial intelligence and machine
learning in finance: A bibliometric review. Research
in International Business and Finance, 61:101646.
Aktas¸, M., Akkus¸ Halepmollası, R., and T
¨
oreyin, B. U.
(2024). Enhancing credit risk assessment with fed-
erated learning through a comparative study. In 8th
EAI International Conference on Robotic Sensor Net-
works.
Assefa, S. A., Dervovic, D., Mahfouz, M., Tillman, R. E.,
Reddy, P., and Veloso, M. (2020). Generating syn-
thetic data in finance: opportunities, challenges and
pitfalls. In Proceedings of the First ACM International
Conference on AI in Finance, pages 1–8.
Awosika, T. et al. (2023). Transparency and privacy: The
role of explainable ai and federated learning in finan-
cial fraud detection. Journal of Financial Technology
and Ethics, 8(1):15–30.
Bottou, L. (2010). Large-scale machine learning with
stochastic gradient descent. In Proceedings of COMP-
STAT’2010, pages 177–186. Springer.
Cao, L. (2022). Ai in finance: challenges, techniques,
and opportunities. ACM Computing Surveys (CSUR),
55(3):1–38.
Cohen, I. G. and Mello, M. M. (2018). Hipaa and pro-
tecting health information in the 21st century. Jama,
320(3):231–232.
Cortes, C. and Vapnik, V. (1995). Support-vector networks.
Machine Learning, 20(3):273–297.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep
Learning. MIT Press.
Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion
probabilistic models. Advances in neural information
processing systems, 33:6840–6851.
Hoofnagle, C. J., Van Der Sloot, B., and Borgesius, F. Z.
(2019). The european union general data protection
regulation: what it is and what it means. Information
& Communications Technology Law, 28(1):65–98.
Hosmer, D. W., Lemeshow, S., and Sturdivant, R. X. (2013).
Applied Logistic Regression. Wiley.
Jolicoeur-Martineau, A. et al. (2023). Generating and im-
puting tabular data via diffusion and flow based gra-
dient boosted trees. Advances in Neural Information
Processing Systems (NeurIPS).
Khaled, A. et al. (2024). Synthetic data generation and
impact analysis of machine learning models for en-
hanced credit card fraud detection. Journal of Artifi-
cial Intelligence and Applications, 12(3):45–60.
Lu, Y., Shen, M., Wang, H., Wang, X., van Rechem, C.,
Fu, T., and Wei, W. (2023). Machine learning for
synthetic data generation: a review. arXiv preprint
arXiv:2302.04062.
McMahan, B., Moore, E., Ramage, D., Hampson, S., and
y Arcas, B. A. (2017). Communication-efficient learn-
ing of deep networks from decentralized data. In Ar-
tificial intelligence and statistics, pages 1273–1282.
PMLR.
Mothukuri, V., Parizi, R. M., Pouriyeh, S., Huang, Y., De-
hghantanha, A., and Srivastava, G. (2021). A survey
on security and privacy of federated learning. Future
Generation Computer Systems, 115:619–640.
Sattarov, E. et al. (2023). Findiff: Diffusion models for
financial tabular data generation. Financial Data Sci-
ence Journal, 9(2):75–90.
Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and
Ganguli, S. (2015). Deep unsupervised learning us-
ing nonequilibrium thermodynamics. In International
conference on machine learning, pages 2256–2265.
PMLR.
Truong, N., Sun, K., Wang, S., Guitton, F., and Guo, Y.
(2021). Privacy preservation in federated learning: An
insightful survey from the gdpr perspective. Comput-
ers & Security, 110:102402.
¨
Ulver, B., Yurto
˘
glu, R. A., Dervis¸o
˘
glu, H., Halepmollası,
R., and Haklıdır, M. (2023). Federated learning in pre-
dicting heart disease. In 2023 31st Signal Processing
and Communications Applications Conference (SIU),
pages 1–4. IEEE.
FEMIB 2025 - 7th International Conference on Finance, Economics, Management and IT Business
88

Wang, H. et al. (2024). A novel federated learning approach
with knowledge transfer for credit scoring. IEEE
Transactions on Neural Networks and Learning Sys-
tems.
Yang, Q., Liu, Y., Chen, T., and Tong, Y. (2019). Federated
machine learning: Concept and applications. ACM
Transactions on Intelligent Systems and Technology
(TIST), 10(2):1–19.
Yeh, I.-C. and Lien, C.-h. (2009). The comparisons of data
mining techniques for the predictive accuracy of prob-
ability of default of credit card clients. Expert systems
with applications, 36(2):2473–2480.
Yurto
˘
glu, R. A., Dervis¸o
˘
glu, H.,
¨
Ulver, B., Halepmollası,
R., and Haklıdır, M. (2024). A novel transformation
through digital twin and federated learning integra-
tion: A case study on cardiovascular disease predic-
tion. In International Conference on Information and
Communication Technologies for Ageing Well and e-
Health, pages 91–113. Springer.
Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., and Gao, Y.
(2021). A survey on federated learning. Knowledge-
Based Systems, 216:106775.
Zhang, L. et al. (2024). The effects of data imbalance under
a federated learning approach for credit risk forecast-
ing. International Journal of Data Mining and Ana-
lytics, 16(4):230–245.
Synthetic Data Generation and Federated Learning as Innovative Solutions for Data Privacy in Finance
89