AI-Driven Fraud Detection in Financial

Transactions with Graph Neural Networks and

Anomaly Detection
2024 International Conference on Science Technology Engineering and Management (ICSTEM) | 979-8-3503-7691-3/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICSTEM61137.2024.10560838

M. Thilagavathi1, R.Saranyadevi2, N. Vijayakumar3, K.Selvi4, L. Anitha5, K. Sudharson6

1
Department of Commerce (A/F), Sir Theagaraya College,Chennai, India.
2
Department of Commerce, A.M.Jain College, Chennai, India.
3
Department of Commerce, B.S.Abdur Rahman Crescent Institute of Science & Technology, Chennai, India.
4,5
Department of MBA, S.A. Engineering College, Chennai, India.
6
Department of AIML, R.M.D. Engineering College, Chennai, India.

E-mail : [email protected], [email protected], [email protected], [email protected]

[email protected], [email protected]

Abstract- Detecting fraudulent financial transactions necessitating innovative approaches to combat financial
is crucial for upholding the integrity of economic crimes effectively. Traditional rule-based and statistical
systems. Traditional methods often lag behind methods often struggle to keep pace with the evolving
evolving fraud tactics, prompting the need for sophistication of fraudulent activities, prompting the
innovative approaches. We propose a pioneering exploration of advanced techniques such as Graph Neural
framework combining Graph Neural Networks Networks (GNNs) and anomaly detection.
(GNNs) with anomaly detection techniques to enhance
fraud detection. Transactions are represented as GNNs, a subset of deep learning models, have gained
graphs, allowing GNNs to capture intricate fraud prominence for their ability to model complex
patterns. Anomaly detection methods flag suspicious relationships and dependencies within graph-structured
transactions. Experimentation on the widely used data [1]. In the context of financial transactions, where
Credit Card Fraud Detection dataset, comprising interactions between entities such as accounts, merchants,
transactions made by European cardholders, and transactions form a network, GNNs offer a promising
showcases substantial advancements over baseline avenue for capturing nuanced patterns indicative of
methods. The dataset is highly unbalanced, with fraudulent behavior [2]. By representing transactions as
fraudulent transactions accounting for only 0.172% of graphs and leveraging GNNs to learn latent representations
all transactions. Our approach achieves a detection of nodes and edges, researchers can uncover hidden fraud
rate of 95% with a false positive rate of 2%, schemes that may evade traditional detection methods.
surpassing the performance of the current state-of-
the-art Gradient Boosting Classifier by 10%. It Anomaly detection techniques complement GNNs by
exhibits resilience against various fraud schemes, identifying deviations from normal transaction patterns,
including account takeover and identity theft. thereby flagging potentially fraudulent activities. With the
Ablation studies underscore the significance of graph- proliferation of large-scale financial datasets, anomaly
based representations and anomaly detection detection algorithms can sift through vast amounts of data
mechanisms. Our research underscores the efficacy of to pinpoint suspicious transactions that warrant further
GNNs and anomaly detection in bolstering financial investigation [3]. These methods play a crucial role in
fraud detection, presenting a promising solution enhancing the effectiveness of fraud detection systems by
against sophisticated fraudulent activities.. reducing false positives and improving the overall
detection accuracy.
Keywords—fraud detection, Graph Neural Networks,
anomaly detection, financial transactions, Credit Card The Credit Card Fraud Detection dataset is one of the most
Fraud Detection dataset extensively researched datasets in the field of financial
fraud detection. This dataset includes transactions that
I. INTRODUCTION were carried out by European cardholders during the
course of two days in September 2013 [4]. The extremely
Detecting fraudulent financial transactions is paramount imbalanced nature of this dataset poses a substantial
for safeguarding economic systems and maintaining public challenge, as fraudulent transactions only account for a
trust in financial institutions. As the digital economy small fraction of the total transactions. This presents a
expands, so too do the methods employed by fraudsters, significant challenge. This kind of class imbalance is a

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.
 significant challenge for conventional machine learning Traditional methods are effective, but they often can't
algorithms, which highlights the requirement for keep up with fraudsters' shifting ideas. Due of this,
specialised methods that are able to properly manage scholars have studied deep learning, graph-based
skewed datasets [5]. algorithms, and ensemble methods to improve fraud
detection accuracy and efficiency.
Deep learning algorithms for financial fraud detection are
a major research area. Convolutional and Recurrent
Neural Networks are examples. These models are perfect
for detecting transactional data fraud because they are
adept at learning complex patterns and correlations in
sequential and structured data. This includes using RNNs
to identify fraud trends in transaction event sequences.

Graph-based methods have also become popular for

financial fraud detection. GNNs are particularly effective
at representing relational data and capturing complex
Fig 1 GNN Model for Financial Transaction. transaction network relationships. Graph neural networks
(GNNs) are neural networks. Researchers discover secret
In recent years, researchers have explored various fraud schemes that may avoid detection. This is done by
strategies to address the challenges posed by imbalanced graphing transactions and using GNNs to learn node and
datasets in fraud detection. Oversampling techniques, such edge latent representations.
as Synthetic Minority Over-sampling Technique
(SMOTE) [6], and cost-sensitive learning approaches have Ensemble approaches like Gradient Boosting Classifier
emerged as popular methods to mitigate class imbalance (GBC) and Random Forest have been popular in fraud
and improve the performance of fraud detection models. detection because they can combine many weak learners
Additionally, ensemble methods, such as Gradient into a powerful classifier. These solutions combine model
Boosting Classifier, have shown promise in achieving high expertise to improve detection and generalisation.
detection rates while minimizing false positives [7]. Ensemble approaches are also resistant to skewed
datasets, making them ideal for detecting rare financial
However, despite these advancements, the detection of transaction fraud.
fraudulent financial transactions remains a challenging and
evolving problem. As fraudsters continue to adapt and Anomaly detection and machine learning help identify
innovate, there is a pressing need for sophisticated, suspicious transactions that deviate from system
adaptive fraud detection systems capable of detecting behaviour. Unsupervised anomaly detection is used to
emerging fraud schemes in real-time [8]. The integration find unexpected transaction patterns. These methods
of GNNs and anomaly detection techniques represents a include Isolation Forest and k-means clustering. These
promising direction in this regard, offering the potential to methods can enhance supervised learning algorithms to
uncover subtle patterns and anomalies indicative of reveal possible fraud.
fraudulent activities that may evade traditional detection
methods [9]. Despite advancements, financial fraud detection faces
many challenges. When faced with skewed datasets with
In this research, we propose a novel framework that a tiny percentage of fraudulent transactions, traditional
combines GNNs with anomaly detection techniques for machine learning techniques struggle. Due to fraud's ever-
robust fraud detection in financial transactions. Leveraging changing nature, detection systems must be able to learn
the Credit Card Fraud Detection dataset as a benchmark, and develop to recognise new schemes.
we aim to demonstrate the effectiveness of our approach in
identifying fraudulent activities with high accuracy and In conclusion, deep learning, graph-based approaches,
minimal false positives [10]. By harnessing the power of ensemble methods, and anomaly detection have advanced
deep learning and anomaly detection, we seek to advance financial fraud detection research. Exploration of these
the state-of-the-art in financial fraud detection and methods powered these advances. By using inventive
contribute to the development of more resilient and methods, researchers want to create more accurate,
adaptive fraud detection systems.. efficient, and adaptive fraud detection systems. These
systems will safeguard economies against financial
II.RELATED WORKS criminality.

Over the past few years, research on better approaches for III.PROPOSED WORK
detecting fraudulent financial transactions has grown.
The representation of transactions as graphs is a powerful

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.
 technique in fraud detection, where each transaction is transformed features V1-V28, which capture the
depicted as a node, and relationships between transactions underlying characteristics of each entity in the transaction
are depicted as edges. This method captures the intricate network. These features serve as input representations for
network of financial interactions, allowing for the the nodes, enabling the graph neural network to learn
detection of anomalous patterns indicative of fraudulent meaningful representations of the entities based on their
activity. By incorporating features such as PCA- transactional attributes.
transformed attributes and temporal information, the graph Edge Features:
model can learn to identify subtle patterns and In addition to node features, edge features are incorporated
dependencies within the transaction data, enabling more into the graph representation to capture temporal
accurate and effective fraud detection. information related to transactions. The temporal
information, derived from the 'Time' attribute, provides
Transaction Data Preprocessing: insights into the timing and sequencing of transactions
Transformation of Numerical Features: within the network. By incorporating edge features, the
The preprocessing stage involves the transformation of graph model can learn to recognize temporal patterns and
numerical features V1-V28 obtained from a principal dependencies in transaction data, further enhancing its
component analysis (PCA) transformation. PCA is predictive capabilities.
employed to reduce the dimensionality of the original
feature space while retaining the essential information
encoded in the data. This transformation helps in capturing
the underlying structure and patterns present in the
transaction data, facilitating more effective modeling and
analysis.
Retention of Non-Transformed Features:
In addition to the PCA-transformed features, the 'Time'
and 'Amount' attributes are retained in their original form.
The 'Time' attribute represents the elapsed time in seconds
between each transaction and the first transaction in the
dataset. This temporal information can be crucial for
identifying patterns and trends in transaction behavior over
time. The 'Amount' attribute denotes the monetary value of
each transaction, providing important context for assessing
the financial impact of fraudulent activities. Fig. 2 GNN Architecture with Anomaly Detection.

Graph Construction: Architecture Selection

Nodes Representation: Choice of GNN Architecture:
Transactions are represented as nodes in the graph, with The selection of the GNN architecture is a critical
each node corresponding to a specific entity involved in decision in the design of the fraud detection model.
the transaction. These entities may include accounts, Options may include Graph Convolutional Networks
merchants, and individual transactions themselves. By (GCNs), which perform convolution operations on graph-
representing entities as nodes, the graph captures the structured data, or Graph Attention Networks (GATs),
interconnected nature of financial transactions, allowing which utilize attention mechanisms to weigh the
for the modeling of complex relationships and importance of neighboring nodes. The choice depends on
dependencies within the data. the specific characteristics of the dataset and the
1. Edges Representation: complexity of the fraud patterns to be detected.
Relationships between entities are represented as edges in Model Complexity and Scalability Considerations:
the graph, indicating transactions or interactions between When selecting a GNN architecture, considerations must
them. For example, an edge between two account nodes be given to the complexity of the model and its scalability
signifies a transaction between the corresponding to large-scale datasets. More complex models may have a
accounts, while an edge between an account node and a higher capacity to capture intricate fraud patterns but may
merchant node represents a purchase or payment also be computationally expensive and prone to
transaction involving the merchant. By encoding overfitting. Therefore, it is essential to strike a balance
transactional relationships as edges, the graph provides a between model complexity and scalability to ensure
comprehensive view of the flow of funds and interactions efficient and effective fraud detection.
between different entities in the financial ecosystem.
Training Procedure
Graph Representation: Input Data Preparation:
Node Features: Before training the GNN model, the input graph data must
The node features in the graph include the PCA- be prepared. This involves encoding the graph structure

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.
 and node features into a format suitable for input into the Semi-supervised Techniques: Semi-supervised anomaly
GNN architecture. Typically, this process involves detection techniques, including Label Propagation and
representing the graph as an adjacency matrix or an edge Self-Training, leverage a small amount of labeled data in
list and encoding node features as feature vectors. conjunction with a larger amount of unlabeled data to
Hyperparameter Tuning: detect anomalies. Label Propagation propagates labels
In the process of training GNN models, there are a number from a small set of labeled examples to unlabeled data
of hyperparameters that constitute an important part of the points based on their similarity, while Self-Training
process. The learning rate (η), batch size, and iteratively updates the model using the most confident
regularisation strength are some of the factors that are predictions on unlabeled data.
considered. To get the highest possible performance on the
validation dataset, hyperparameter tuning involves
Threshold Selection:
tweaking these parameters to achieve the best possible
results. This is done in order to achieve the best possible Determination of Anomaly Score Threshold: Anomaly
results. Approaches such as grid search and random search detection algorithms typically assign anomaly scores to
are examples of strategies that can be utilised in order to each data point, indicating its likelihood of being
efficiently explore the hyperparameter space. Other anomalous. The threshold for classifying a data point as
techniques for this purpose include random search. anomalous is determined based on these anomaly scores.
Training Process: This threshold is often set empirically or using domain
In order to minimise a loss function L, which quantifies knowledge, aiming to balance the trade-off between the
the difference between the predicted and true labels of the detection rate (sensitivity) and false positive rate
training data, the parameters of the GNN model are (specificity).
modified throughout the training phase. This is done in
order to keep the loss function as small as possible. To Trade-off between Detection Rate and False Positive Rate:
The selection of the anomaly score threshold involves a
accomplish the objective of reducing the loss function to
the smallest possible value, this is done. As a result of this, trade-off between the detection rate and false positive rate.
it is feasible to bring the loss function down to its smallest A lower threshold increases the detection rate by capturing
more anomalies but also leads to a higher false positive
possible position. Common optimisation algorithms that
are used in the process of training GNNs include the rate by classifying more non-anomalous data points as
stochastic gradient descent (SGD) algorithm, which anomalies. Conversely, a higher threshold reduces the
false positive rate but may result in missed anomalies.
updates model parameters based on the gradient of the loss
function with respect to each parameter, and the Adam Achieving an optimal balance between these two metrics
optimizer, which adjusts the learning rate for each is crucial for effective anomaly detection in fraud
parameter based on previous gradients and squared detection systems.
gradients. Both of these algorithms are used in the process
of training GNNs. When it comes to the process of training IV.EXPERIMENTATION SETUP:
GNNs, both of these methods are being utilised. These two
approaches are employed in order to achieve the goal of Dataset Description:
achieving the highest possible rate of learning. The 'Time' and 'Amount' characteristics, which indicate
the seconds that pass between each transaction and the
Based on the graph data that is provided, the GNN model dataset's initial transaction and the transaction amount,
is able to learn how to minimise the loss function and respectively, are still present in the dataset.
accurately forecast fraudulent transactions. This is
accomplished by learning how to learn. Gradient descent, Issues with Class Imbalance: The class imbalance
which is employed to repeatedly update the parameters, is problem in the dataset—fraudulent transactions make up
the method that is utilised in order to accomplish this goal. just 0.172% of all transactions—is one of its major
problems. Because there are so few positive examples,
Integration of Anomaly Detection Techniques conventional machine learning algorithms may find it
Anomaly Detection Algorithms: difficult to correctly identify fraudulent transactions as a
Unsupervised Techniques: Unsupervised anomaly result of this extreme class imbalance.
detection methods, such as Isolation Forest and k-means
clustering, are utilized to identify anomalies in the Preprocessing Steps:
transaction data without the need for labeled training Data Cleaning: Addressing outliers, missing numbers, and
examples. Isolation Forest operates by isolating anomalies other problems with the quality of the data is part of the
in sparse regions of the feature space using random preparation phase. Data cleaning may concentrate on
partitioning, while k-means clustering partitions the data removing outliers that have the potential to distort the
into clusters and identifies outliers based on their distance results because the dataset only includes numerical
from the cluster centroids. features that came about as a result of PCA
transformation.

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.
 Train-Test Split: To assess the efficacy of the fraud techniques like GNNs in capturing the intricate patterns
detection model, the dataset is split into subsets for inherent in financial transaction data, highlighting the
training and testing. In order to guarantee that the training potential for significant improvements in fraud detection
and testing datasets retain the same class distribution, a systems.
random or stratified split is usually utilised. The testing
set is used to assess the model's performance on untested
data, whereas the training set is used to train the model.

Model Evaluation Metrics:

Detection Rate: The percentage of accurately identified
fraudulent transactions among all real fraudulent
transactions is measured by the detection rate, which is
sometimes referred to as the sensitivity or true positive
rate.

V.RESULTS AND DISCUSSIONS

In the following section, a comprehensive examination of
the findings presented in the table is presented. This Fig. 3 Detection Rate.
analysis emphasises the benefits and limitations of each
model, as well as the enhanced performance of the False Positive Analysis:
suggested strategy in detecting fraudulent transactions. The detailed analysis of false positives (FP) reveals
Additionally, this analysis highlights the limitations of significant variations in the performance of different
each respective model. models. While the proposed GNN + Anomaly Detection
model demonstrates the lowest count of false positives and
Detection Rate: the lowest false positive rate (FPR), Gradient Boosting
In order to gain significant insights into the effectiveness Classifier and Random Forest exhibit higher counts and
of each strategy in recognising fraudulent transactions, it rates of false positives. These findings underscore the
is necessary to compare the detection rates of various effectiveness of the proposed model in minimizing false
models. The model displays improved performance in alarms while accurately detecting fraudulent transactions.
comparison to typical machine learning models, which
highlights its potential as a promising solution for TABLE 2 FPR COMPARISON
enhancing fraud detection systems in financial areas such Fraudul
Total Non-
ent
as banking and finance. For the purpose of further Model Transacti Fraudulent FP FPR
Transac
improving the model's performance and scalability in ons Transaction
tions
real-world applications, subsequent study could
Proposed
investigate the possibility of implementing additional 2,84,807 492 2,84,315 5 2%
Model
additions and optimisations.
GBC 2,84,807 492 2,84,315 14 5%
Table 1 DETECTION RATE COMPARISON RF 2,84,807 492 2,84,315 20 7%
Model Detection Rate
Proposed GNN + Anomaly Detection 95% IF 2,84,807 492 2,84,315 9 3%
Gradient Boosting Classifier (GBC) 85%
Random Forest (RF) 78%
The comparison highlights the significant improvements
Isolation Forest (IF) 92%
offered by the proposed GNN + Anomaly Detection model
over the current state-of-the-art Gradient Boosting
The comparison reveals the proposed model's notable
Classifier. With a 10% increase in detection rate and a
superiority, achieving a detection rate of 95% compared to
lower false positive rate, the proposed model demonstrates
the baseline models. This underscores the effectiveness of
superior performance in detecting fraudulent transactions
integrating Graph Neural Networks (GNNs) with anomaly
while minimizing false alarms. These findings underscore
detection techniques for fraud detection. While Gradient
the effectiveness of leveraging Graph Neural Networks
Boosting Classifier and Random Forest show respectable
(GNNs) with anomaly detection techniques in enhancing
detection rates of 85% and 78% respectively, they fall
fraud detection systems and advancing the state-of-the-art
short of the proposed model's performance. The Isolation
in financial security.
Forest, with a detection rate of 92%, performs
competitively but doesn't surpass the proposed model.
Overall, the results underscore the importance of advanced

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.
 involves refining feature engineering for nuanced
transaction pattern capture, advancing Graph Neural
Network architectures for enhanced fraud detection
accuracy, developing dynamic learning algorithms for
real-time adaptation to evolving fraud tactics, and
prioritizing interpretability for stakeholder trust and
understanding. These endeavors will fortify fraud
detection systems, ensuring economic integrity and
financial security.

REFERENCES
[1] Khemani, B., Patil, S., Kotecha, K. et al. A review of graph
neural networks: concepts, architectures, techniques, challenges,
datasets, applications, and future directions. J Big Data 11, 18
(2024). https://doi.org/10.1186/s40537-023-00876-4
Fig. 4 FPR Comparison. [2] Wei, S.; Lee, S. Financial Anti-Fraud Based on Dual-Channel
Graph Attention Network. J. Theor. Appl. Electron. Commer.
Res. 2024, 19, 297-314. https://doi.org/10.3390/jtaer19010016
VI.CONCLUSION AND FUTURE WORKS [3] Hilal, W., Gadsden, S. A., &Yawney, J. (2022, May). Financial
Fraud: A Review of Anomaly Detection Techniques and Recent
The research introduces an innovative framework that Advances. Expert Systems With Applications, 193, 116429.
https://doi.org/10.1016/j.eswa.2021.11642
merges Graph Neural Networks (GNNs) with anomaly [4] G. R et al, "An Effective Copyright Management Principle
detection methods to bolster fraud detection in financial using Intelligent Wavelet Transformation based Water marking
transactions. Experimentation on the Credit Card Fraud Scheme," (ACCAI), Chennai, India, 2022, pp. 1-7, doi:
Detection dataset demonstrates promising results, with the 10.1109/ACCAI53970.2022.9752516.
[5] R. T.et al, "COVID-19 Identification and Detection from CT-
proposed model achieving a remarkable 95% detection Images using AI Based Ensemble Model," (ICSES), Chennai,
rate, surpassing traditional methods like Gradient Boosting India, 2022, pp. 1-10, doi: 10.1109/ICSES55317.2022.9914227.
Classifier and Random Forest. Moreover, the model [6] Dornadula, V. N., & Geetha, S. (2019). Credit Card Fraud
maintains a low false positive rate of 2%, indicating its Detection using Machine Learning Algorithms. Procedia
Computer Science, 165, 631–641.
efficacy in minimizing false alarms. These findings https://doi.org/10.1016/j.procs.2020.01.057
underscore the potency of GNNs and anomaly detection [7] Manda, V. T., Kondapalli, D., Malla, A. S., M, J. N., &Charan,
techniques in elevating fraud detection capabilities and Y. (2024, February 20). Imbalanced Data Challenges and Their
advancing financial security standards. Future work Resolution to Improve Fraud Detection in Credit Card
Transactions. https://doi.org/10.21203/rs.3.rs-3962043/v1
involves refining feature engineering for nuanced [8] Du H, Lv L, Wang H, Guo A (2024) A novel method for
transaction pattern capture, advancing GNN architectures detecting credit card fraud problems. PLOS ONE 19(3):
for enhanced fraud detection accuracy, developing e0294537. https://doi.org/10.1371/journal.pone.0294537
dynamic learning algorithms for real-time adaptation to [9] Gururaj, N., Vinod, V. & Vijayakumar, K. Deep grading of
mangoes using Convolutional Neural Network and Computer
evolving fraud tactics, and prioritizing interpretability for Vision. Multimed Tools Appl (2022).
stakeholder trust and understanding. https://doi.org/10.1007/s11042-021-11616-2.
[10] A. G, et al, "Design of Miniaturized Single Bit MEMS Phase
By addressing these areas, future research can contribute Shifter using MEMS Switches," (ICOEI), Tirunelveli, India,
significantly to the development of robust, efficient, and 2021, pp. 235-239, doi: 10.1109/ICOEI51242.2021.9453063.
trustworthy fraud detection systems, safeguarding
economic integrity and financial well-being. Future work

Authorized licensed use limited to: New Horizon College of Engineering - Bengaluru. Downloaded on November 09,2024 at 04:10:44 UTC from IEEE Xplore. Restrictions apply.