Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)

IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

PaySafe Al: Intelligent Fraud Detection for UPI

Transactions using Machine Learning
2025 International Conference on Intelligent Computing and Control Systems (ICICCS) | 979-8-3315-1208-8/25/$31.00 ©2025 IEEE | DOI: 10.1109/ICICCS65191.2025.10985175

Markala Anjali 1, Maddela Ajay2, Mukkerla Saiteja3, Baddula Omkar Yadav4

1Assistant
Professor 2,3,4UG Scholar
1,2,3,4Department
of Computer Science and Engineering (AI & ML)
Vignan Institute of Technology and Science
Nalgonda ,Telangana-508284
1
[Link]@[Link], 2ajaymaddela0906@[Link] , 3saitejacareeracademy@gmail,com, 4omkary4273@[Link]

Abstract—In the financial world, it is impoartant to have a B. Online Card Purchases:

sound understand of systems against fraud, as they help detect
activities such as anomalous and malicious transactions, through Online Card transactions or telephonic purchases make use of this
the analysis of activity engagement. Decision Tree are the recent feature more often. In these instances, the customer only requires
form of deep learning models which, in more recent years, have providing a select few key card details, which includes the card
shown advanced capability in the identification of fraudulent
number, its expiration date, and its secure code also known as CVV.
transactions, however, sensitivity and accuracy when
implemented in real time systems remain an issue. In this When it comes to unauthorized transactions in this context, the
research, we provide a new idea of fraud detection of UPI physical card is not required, but rather these unique card details are
transactions that utilizes a Decision Tree based sequence required to be illicitly acquired. This information can be acquired
modeling framework. The model focuses on normalization sets through a number of ways such as phishing, data leaking and
and typical cardholder activities to effectively learn how to detect
malware. The main issue with virtual card fraud is the fact that the
anomalies in UPI transactions. It then analyzes the sequences of
a person’s activities and adds on the probabilities, to then actual owner of the card may not realize that their information has
classify if the transaction is fraudulent or not. Several been hacked as the fraudulent user does not need to show the card.
measures are established, however, in order to aid in This type of fraud becomes quite alarming because it is possible for a
minimizing false positives, because the aim is to ensure valid perpetrator to take advantage of a transaction without detection for a
transactions are still processed. The experimental results support
long period, resulting in massive losses economically from a
the proposed fraud detection system’s potential usefulness in
monitoring actual UPI transactions and it also shows how it cardholder’s account Fraud.
outperforms current systems in related studies. Our model
facilitates the UPI transaction environment equally and C. Fraud Detection Using Behavior Approach:
addresses the fine balance in detection efficiency, accuracy and One of the techniques that seem to be very effective in locating
transaction speed by proposing the integration of deep learning
fraud where the method of payment is a UPI and especially when
techniques.
Keywords— Fraud Detection, UPI Transactions, Decision Tree , purchasing by using virtual cards is assess the cardholders’ adopting
Deep Learning, Sequential Modeling, Financial Security,Logistic spending behavior. Every individual possesses a fairly distinctive set
Regression,SVM. of behaviors towards ‘how and when’ purchase their items. These
operating behaviors can be classified into some broad determinants
I. INTRODUCTION which include the following and so forth: Purchase Category: The
products or services purchased by the cardholder in the normal
In the past several years, online shopping has progressed at
course of business Transaction Frequency: The amount of
a dizzying pace due to the development of technology, ongoing
transactions made by the cardholder, whether it is on a daily basis,
expansion of the internet, and ease for consumers. In turn, the
weekly, or on a monthly scale.
number of online shoppers and online transactions has grown
in the whole world. D. Fraud Detection Methods:
A. Physical Card Purchases: From the definition, the detection of such discrepancies is of great
concern hence a number of methods have been set and used to
In this method of transaction, the merchant is presented the
enhance fraud detection in UPI transactions. These methods
card in order to complete the payment so with the use of a
generally fall under two categories:
card reader to process the payment, the transaction is face
1) Rule-based Systems:: These are systems which have rules
to face. In this scenario, the loss of the card is a significant
established beforehand on the spending behavior of a customer and
factor aiding in the fraudulent activity. If the original owner
the behavior is deemed to be out of the ordinary use of the customer.
loses possession of the card, for whatever reason, then the
For instance, rules can be enacted such that if a specified dollar value
buyer can easily gain access to funds that don’t belong to
is exceeded, an alert is triggered as with purchases made in countries
them, which results in losses for both the credible cardholder
different from what the card holders have previously done. Despite
and the company facilitating the UPI transaction. It can be
being simple to implement, these systems have limitations especially
said that in case the claimed loss has been hidden for a longer
in terms of their application against more complex patterns of fraud.
duration, then the loss can be even greater since concealing
the loss allows the owner to cause more damage.

Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025
979-8-3315-1208-8/25/$31.00 ©2025atIEEE
[Link] UTC from IEEE Xplore. Restrictions apply.1557
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

Deep learning methods have been increasingly applied to

II. RELATED WORKS fraud detection. Karthick et al. [8] introduced a deep
Financial fraud has emerged as a developing issue in regression network for multimedia text image super-
the digital age, requiring efficient detection tools. The resolution, which has the potential to improve fraud detection
incorporation of artificial intelligence (AI), machine in document verification tasks. Bagga et al. [11] used
learning (ML), blockchain, and data mining ensemble learning in credit card fraud detection to enhance
methodologies has been promising against fraud. This classification accuracy. Fanai et al. [12] proposed a deep
literature review reviews current research on financial autoencoder with deep classifiers for detecting fraud, showing
fraud detection, cyber threat trends, and AI-based better performance in detecting fraudulent transactions.
solutions. Benchaji et al. [13] improved fraud detection through
attention mechanisms and LSTM deep models, showing better
accuracy in financial datasets. Cheah et al. [14] handled class
Sharif et al. [1] presented a broad overview of cyber
imbalance in fraud detection with a hybrid SMOTE-GAN
threats and their financial implications, with the
approach to improve fraud detection accuracy. Nami et al.
growing sophistication of cyberattacks and their
[15] proposed a cost-sensitive dynamic random forest and k-
financial effects being the key point. The research
nearest neighbors-based method for fraud detection to
outlined emerging trends in cybersecurity in the future,
minimize false positives and enhance detection efficiency.
such as defense mechanisms using AI and the
implementation of blockchain. Bao et al. [2] examined
the use of AI for detecting fraud and recognized its role The literature emphasizes the development of fraud
in pattern identification, anomaly detection, and risk detection techniques, with a focus on AI, deep learning, and
analysis. The research highlighted AI in decision- blockchain as the major technologies. Future research needs to
making for finances, fraud detection, and process concentrate on increasing model interpretability, lowering
efficiency. Orioha et al. [3] suggested the use of computational expenses, and combining multiple detection
blockchain technology in the detection of counterfeit methods to increase financial security.
products, where its strengths were highlighted in
guaranteeing data integrity and transparency. III. METHODOLOGY
Blockchain application is also extended to fraud
detection in financial transactions to minimize data A. Working
tampering and unauthorized access. Al-Hashedi et al.
[4] discussed methodologies for the detection of
financial fraud through data mining from 2009 to 2019. 1) Input Data Processingn: Transaction Input: In other
The research classified fraud detection methods words, each UPI transaction comes in with certain features and
according to supervised and unsupervised learning is rendered as a separate data point: Identifying number of the
approaches, pointing to outstanding improvements in transaction User identification number Value of the transaction
classification algorithms and the accuracy of fraud Date and time the transaction was carried out User location
detection. Previous actions of the user. Processing: Standardize numeric
data (i.e. amount transacted). Transform categorical attributes
Supervised learning has been used extensively in (place, payment type, etc.) into numeric values. To specify the
fraud detection. Osisanwo et al. [5] contrasted input, construct a transaction matrix.
classification algorithms, proving to be effective for
detecting fraudulent transactions. In a similar fashion, 2) Feature Representation: Each transaction and transac-
Vlachas et al. [6] analyzed backpropagation and tions that took place before it can be expressed in a two
reservoir computing to predict financial anomalies. dimensional matrix. Rows: Individual features such as amount,
Thudumu et al. [7] surveyed anomaly detection time and location. Columns: A series of past transactions of
methods in high-dimensional data, presenting some a particular user. This forms the input of the CNN model and
statistical and machine learning-based methods. The hence it is possible for it to detect patterns across space and
paper emphasized the role of feature selection and time.
dimension reduction in detecting financial fraud.

979-8-3315-1208-8/25/$31.00 ©2025 IEEE 1558

Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

3) Classification Decision: This is where a particular cut- Algorithm 1 Model Processing Algorithm for UPI Fraud
off has been set for example above point 7 and whose point Detection
defines fraud detection limits for the attributes. If the model Require Preprocessed numerical data (tensor format)
output exceeds the figure set in place, then the transaction can Ensure Fraud detection results from each model
be identified as being of a Fraudulent nature, otherwise it can 1) Step 1: Extract features using Random Forest
be marked as being Genuine. 1.1 Feed the preprocessed data tensor into the Random
Forest model.
B. Algorithms 1.2 Random Forest extracts features and builds deci-
Preprocessing Algorithms: Data Normalization Translates sion trees based on input data.
numerical data such as transaction amount to a standard 1.3 Use the models to predict the probability of fraud
range of [0,1] or [-1,1] Algorithm: Min-Max Scaling or Z- for the each transactions.
Score Normalization normalized = min() max() min() 2) Step 2: Extract features using Logistic Regression
x normalized =max(x)min(x)xmin(x) Categorical Encoding 2.1 Feed the preprocessed data tensor into theLogistic
Transforms categorical features such as location or payment Regression
type into respective numeric values. Algorithms: One-Hot 2.2 Logistic Regression processes the features and
Encoding: Transforms every category into binary vectors. gen- erates a fraud prediction score
Label Encoding: Replaces categories with distinctive numbers. 3) Step 3: Extract features using Support Vector Machine
Transaction Matrix Construction Gathers the previous and (SVM)
present transaction history in a single format ready for CNN 3.1 Feed the preprocessed data tensor into the SVM
input Row: Features( e.g. Amount, Location, Time) Column: model
Sequence of Transactions Ensemble Technique: The use of 3.2 SVM identifies support vectors and classifies trans-
several models enhances the power of the algorithm, which actions as fraudulent or legitimate
minimizes the chances of misclassification occurring. 4) Step 4: Aggregate results from all models
Feature Variety: Separate models may use different methods
4.1 Combine predictions from Random Forest,Logistic
for feature selection and feature engineering, which adds
Regression, and SVM
further value to the analysis.
4.2 Apply a voting mechanism or averaging to deter-
This algorithm describes a model processing workflow
mine final fraud prediction
for fraud detection within Unified Payments Interface (UPI)
payments by using 3 ML models: Random Forest, Gradient 5) Step 5: Return the aggregated fraud detection results
Boosting, and support vector machine. The models in question 5.1 Fraud detection results generated by Random For-
will analyze preprocessed data to predict if a transaction has est,Logistic Regression, and SVM
the potential to be fraudulent. The explanation is as follows
step by step.
algorithm space separating directions of lawful and fraudulent financial
transactions.
1) Step 1: Using Random Forest to Extract Features: Input
4) Step 4 encompasses Aggregation of the results from all
Data: The preprocessed data is provided in tensor form, which
models: This step involves combining the predictions made
is a structured representation of the transaction information
by each of the models. In our case, the aforementioned models
in a numeric format. Feature extraction: The Random Forest
based on random forest and support vector are combined with
model examines the input data, and constructs a set of decision
the ones based on gradient boosting. Based on the constructs
trees to extract features. The data from each tensor is used to
of the models, fraud votes are offered by all independent
split the target’s trees, according to importance of features.
models. Therefore, after obtaining the final results, votes from
Fraud Probability Prediction: It indicates the cash outflows in
all models are aggregated and the result with higher fraud votes
which the model predicts each transaction with a probability
is selected. Alternatively, scores from fraud votes coming from
of a specific amount of fraud that can occur.
all three models can be averaged and a single integrated fraud
2) Step 2: Using Logistic Regression To Extract Features:
score can be obtained.
Input Data: The same preprocessed data tensor is fed into a
5) Step 5 Involves Return Aggregated Fraud Detection
Gradient Boosting model. Processing: The Gradient Boosting
Results: Aggregated results offer a strong suggestion of
algorithm improves an input progressively over a sequence of
fraudulent activities and therefore reinforce the claim of higher
weak models .
confidence. Achievement of these results is based on the
3) Step3 feature extraction using the Support Vector Ma- amalgamated strength residing within the random forest,
chine (SVM):: which corresponds directly to the input data. Logistic Regression, and SVM fraud detection models.
SVM support vector model is included in the classification
phase and uses a tensor with previously categorized data C. proposed system
points. The classification phase covers the construction of the The Identifying Consumer Spending System (ICSS) pro-
SVM-based classifier, which captures the super hyperplane in vides a solution in the form of a unique architecture. This

979-8-3315-1208-8/25/$31.00 ©2025 IEEE 1559

Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

architecture has two specific tasks to carry out. Firstly, it is able

to analyze security protocols by distinguishing between good
and bad users. Secondly, it automatically adjusts the limits on
transactions that are being carried out. This reaction would be
based on the user’s normal spending range and any variations
present within it. The pre-processed transaction parameters
that are previously mentioned would be fed into the Decision
Tree. As more information gets stored with the transactions, the
general Spending and Purchasing Range (SPR) of the user
would also change and so will Decision Tree. input
information. Hence, the Decision Tree. would be able to
categorize a transaction as either being genuine or upset due to
the variance in spending range. The anomaly based intrusion
detection algorithm accomplishes this task by interpolating
and extrapolating transactions with respect to the users’
Fig. 1. System Architecture
behavioral patterns in relation to spending and purchasing.
The purchasing range covers all sorts of activities that include
sharing a ride or buying a vehicle. After analysing various
customer behavior ranges and covering a span of over 8 prototypes or mockups of the desired system’s functionality
months, the radical alteration in purchasing behavior due to and interface are created. The users play an active role in
investment or business activity was detected. Further this phase to get feedback for refining the designs to meet
investigations led up to discovering a bundle of cash their expectations. Following this approval of the design, the
withdrawal activities that resulted in losses from the degree process shifts to the Construction phase, where the system
of is built, tested, in short cycles. The developers and the users
$1100 to $1400 every day. This was detected in the case of work together and adjust many times continuously due to real-
new or young investors making the transition into investments time feedback from ongoing use. Thus, this loop of design and
activity after a boasting economic activity. construction will ensure continuous refinement and flexibility.
Jessica produces an academic paper that explores how Finally, it is in the Cutover phase, where the system goes
Decision Tree and other algorithms protect users and financial to production after the implementation, user training, and
institutions from fraud during UPI transactions. Her the transition from the old system to the new system. This
approach is based on reliable mathematical models and approach stresses user involvement, flexibility, and iterative
leverages the theoretical work of her MSc supervisor to refinement so that the final system effectively meets both
substantiate her claims. functional and user requirements.
The Incremental Development Model is widely used in
D. System Architecture software engineering, especially where the old methodologies
This diagram presents a process flow on managing user are inefficient to develop by fragmenting the system into
interactions in a system, which might include an account small-sized, easily workable units called increments. After
creation process, login process, transaction-making process, a general analysis and designing stage of the entire system,
and the security-verification process. Below is the step-by- there is the development of a general scope, objectives, and
step description: New Account In the beginning, the user overall architecture. This high-level planning ensures cohesion
starts the process with their User Info in order to create a in the overall structure of the system while setting up stages
account. Login After a new account has been created, the for incremental development. Once all the planning has been
users log into his account with Login Info. Transaction, when done at the top, the system gets divided into its functional
accessed, using a Login, the user may initiate a Transaction increments with independent analysis, design, implementation,
by filling up details of a transaction Verifications before the and testing processes. The prioritization is again based on
actual processing of the transaction, the classification occurs. stakeholders’ needs for delivering critical or value-for-money
The process of verification communicates with the Security features earlier in the development timeline.
system by exchanging Security Info to authenticate the A key feature of this model is the parallel or overlapping
transaction as being safe and authorized Complete Transaction development of increments, as shown in the diagram. In this
After being verified successfully, the Verification Details is model, while one increment is in the coding phase, subsequent
sent to complete the transaction. increments may already be undergoing analysis or design. This
Figure 1 shows a phased and iterative approach to system overlap minimizes idle time, speeds up delivery, and brings
development, which is commonly associated with about a faster realization of value. Once an increment is fully
methodologies such as Rapid Application Development tested, it gets deployed, thus providing a working part of
(RAD). The process starts at the Requirements Planning phase the system that may be used, reviewed, and evaluated by the
where all the stakeholders collaborate to define the goals, stakeholders. Feedback from these reviews is then assimilated
scope, and key requirements of the project. This ensures that
user needs are aligned to the outcomes expected in the system.
This is followed by the User Design phase, where interactive
Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025
979-8-3315-1208-8/25/$31.00 ©2025atIEEE
[Link] UTC from IEEE Xplore. Restrictions apply.1560
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

into the future increments so that continuous improvement and

alignment with user expectations are achieved.
This approach is highly flexible and adaptive, making it
appropriate for projects whose requirements are in flux. The
Incremental Development Model reduces risks by delivering
small, tested units of functionality at regular intervals, thereby
identifying and resolving potential issues early in the devel-
opment process. It also improves stakeholder satisfaction by
allowing them to interact with and influence the system as
it evolves. It also enables better resource management by
allowing teams to focus on smaller, well-defined tasks instead
of trying to deliver the whole system in one go. Therefore,
the Incremental Model gives a balance between structure and
adaptability, so that projects can be delivered efficiently and
of high quality.
IV. RESULT AND DISCUSSION
Fig. 3. ROC curve of Logistic Regression
The Machine Learning-Based Intrusion Detection System
is conceived with the objective of effectively detecting cy-
berattacks by leveraging several modules used for handling
the collection of data, extraction of features,
of anomalies, classification, and visualization ofidentification
the results.
Every module in these systems contributes substantially to
achieving precise and accurate detection of fraud attempts in
financial transactions.
In this, a wide range of efficient approaches are included
for detecting the diabetes mellitus with high accuracy and
performance metrics
Support Vector Machine (SVM) :
SVM was introduced by Cortes, C., and Vapnik, V. (1995)
[9] is a powerful classifier and regression algorithm, which is
effective for high-dimensional spaces. This method constructs
a hyper-plane it attempts to fit as many instances as possible
within a margin of tolerance, making it useful for accurate Fig. 4. ROC curve for Support Vector Machine
prediction of solar irradiance
TABLE I
Random Forest Algorithm : LOGISTIC REGRESSION SUMMAARY

The random forest algorithm, a popular machine learning Estimate Std. Error z value Pr(¿—z—)
technique, was introduced by Leo Breiman in 2001[10] . It is (Intercept) -4.253e-01 5.127e-01 -0.830 0.40676
Id 1.068e-03 4.161e-04 2.567 0.01026
a supervised learning technique in Ensemble Machine
transhour -1.364e-02 1.135e-02 -1.202 0.22925
Learning to make predictions. transday -3.350e-03 1.177e-02 -0.285 0.77597
transday -1.281e-02 3.729e-02 -0.344 0.73112

Furthermore, the system was implemented to support real-

time transaction processing, so it can process an adequate
amount of transactions in a smooth manner. This is particularly
crucial for UPI systems, where millions of transactions are
happening on a daily basis, and fraud detection in real-time
is of vital importance to avoid financial losses. The addi-
tion of a feedback loop mechanism* enhanced the system’s
ability to adapt, enabling it to learn from changing fraud
methods through ongoing retraining on new information. This
guaranteed that the model continued to work effectively as

Fig. 2. DECISION TREE using R program

Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
979-8-3315-1208-8/25/$31.00 ©2025 IEEE 1561
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

TABLE III
TABLE II SUMMARY OF RESULT
COMPARING ACCURACY OF ALGORITHMS
[Link] CP nsplit rel error xerror xstd
Algorithm Accuracy [1] 0.5327 01.0000 1.0000 0.0420
[2] 0.08974 10.46723 0.47863 0.03340
Support 85.9%
[3] 0.03418 30.28774 0.33618 0.02890
Vector
[4] 0.02706 40.25356 0.24501 0.02516
Machine
[5] 0.01994 60.19943 0.24786 0.02529
Logistic 84%
Regres-
sion curve analysis was used to con- firm the performance of
Random 84% the system. An elevated ROC-AUC score validated that the
Forest
model had a robust balance between sensitivity (identifying
genuine fraud cases) and specificity (preventing false
alarms). This measure illustrated that the system was able to
effectively differentiate between fraudulent and authentic
transactions, rendering it a robust solution for fraud risk
detection. On the whole, the integration of high accuracy,
real-time computation, flexibility, and robust ROC
performance makes this model a highly stable and efficient
solution to fraud detection.
V. CONCLUSION

By integrating a SVM into the fraud risk classification

analysis, there was a remarkable improvement in predictive
performance compared to Decision Tree machine learning
models. Although the Random Forest and Logistic Regression
did moderately better with an accuracy of 84.9%, the SVM
model performed better than all, with a significantly higher
Fig. 5. ROC curve for Random Forest accuracy of 85.9%. The ROC curve for the SVM showed a
sharp rise towards the top-left quadrant of the graph, which
meant a high true positive rate along with a low false positive
rate, which is a sign of better classification ability. This
implies that the SVM was able to extract intricate, non-linear
patterns in the data that standard models were incapable of.
Moreover, the lowest performance was recorded by the
Decision Tree algorithm further indicating the superiority of
deep learning methods in fraud detection. In general, the
findings point out that the application of SVM significantly
improves the accuracy and reliability of fraud risk prediction,
and thus it is a promising method for future use in financial
fraud detection systems.
REFERENCES
Fig. 6. Results
[1] M. H. U. Sharif and M. A. Mohammed, “A literature review of
financial losses statistics for cyber security and future trend,” IEEE
Access, vol. 15, pp. 138–156, 2022.
fraudulent methods changed with time. Compared to other [2] Y. Bao, G. Hilary, and B. Ke, “Artificial intelligence and fraud
detection,” in Innovative Technology at the Interface of Finance and
machine learning methods like Random Forest, SVM, and Operations (Springer Series in Supply Chain Management, Forth
Logistic Regression, the Decision tree-based system was coming), vol. 1. Springer, 2022,
much better at identifying intricate fraud patterns, particularly [3] Orioha, Ebuka. "Fake Product Detection with Blockchain
Technology." Journal of Artificial Intelligence and Capsule Networks
in extremely imbalanced datasets where fraudulent 6, no. 3 (2024): 316-331.
transactions are few. Although Decision Trees are helpful for [4] [Link]-Hashedi and P. Magalingam, “Financial fraud detection
interpretability and the identification of most important risk applying data mining techniques: A comprehensive review from
2009 to 2019,” IEEE Access, vol. 40, 2021, Art. no. 100402.
factors, they tend to perform poorly with imbalanced data, [5] F. Y. Osisanwo, J. E. T. Akinsola, O. Awodele, J. O.
resulting in lower detection accuracy. Yet, Decision Trees are Hinmikaiye,
valuable in an-alyzing transaction behaviors and providing the O. Olakanmi, and J. Akinjobi, “Supervised machine learning algo-
rithms: Classification and comparison,” Int. J. Comput. Trends
basis for more complex ensemble models. In addition, ROC Technol. (IJCTT), vol. 48, no. 3, pp. 128–138, 2017.
[6] P. R. Vlachas, J. Pathak, B. R. Hunt, T. P. Sapsis, M. Girvan,E. Ott, and
P. Koumoutsakos, “Backpropagation algorithms and reservoir
computing in recurrent neural networks for the forecasting of com
plex spatiotemporal dynamics,” Neural Netw., vol. 126, pp. 191–
217, Jun. 2020.
Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
979-8-3315-1208-8/25/$31.00 ©2025 IEEE 1562
 Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS-2025)
IEEE Xplore Part Number: CFP25K74-ART; ISBN: 979-8-3315-1208-8

[15] S. Nami and M. Shajari, “Cost-sensitive payment card fraud detection

[7] S. Thudumu, P. Branch, J. Jin, and J. Singh, “A comprehensive survey of based on dynamic random forest and k-nearest neighbors,” Nov. 2018.
anomaly detection techniques for high dimensional big data,” J. Big
Data, vol. 7, no. 1, pp. 1–30, Dec. 2020.
[8] Karthick.S and Muthukumaran.N, "Deep Regression Network for the
Single Image Super Resolution of Multimedia Text Image," 2023 IEEE
5th International Conference on Cybernetics, Cognition and Machine
Learning Applications (ICCCMLA), Hamburg, Germany, 2023, pp. 394-
399, doi: 10.1109/ICCCMLA58983.2023.10346975
[9] Cortes, C., and Vapnik, V. (1995). Support-vector networks. Machine
Learning, 20(3), 273–297.
[10] L. Breiman. Random forests. Machine Learning, volume 45, pages 5–32,
2001.
[11] S. Bagga, A. Goyal, N. Gupta, and A. Goyal, “Credit card fraud
detection using pipeling and ensemble learning,” Proc. Comput. Sci.,
vol. 173, pp. 104–112, Jan. 2020.
[12] H. Fanai and H. Abbasimehr, “A novel combined approach based on
deep autoencoder and deep classifiers for credit card fraud detection,”
Expert Syst. Appl., vol. 217, May 2023, Art. no. 119562.
[13] I. Benchaji, S. Douzi, B. El Ouahidi, and J. Jaafari, “Enhanced credit
card fraud detection based on attention mechanism and LSTM deep
model,” J. Big Data, vol. 8, no. 1, pp. 1–21, Dec. 2021.
[14] P. C. Y. Cheah, Y. Yang, and B. G. Lee, “Enhancing financial fraud
detection through addressing class imbalance using hybrid SMOTE-
GAN techniques,” Int. J. Financial Stud., vol. 11, no. 3, p. 110, Sep.
2023.

979-8-3315-1208-8/25/$31.00 ©2025 IEEE 1563

Authorized licensed use limited to: Zhejiang University. Downloaded on August 24,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.