Online financial fraud leads to billions of dollars losses and acute challenges across the world due to pervasive threat on digital adoption in banking, financial and Insurance (BFSI) sectors. It outpaces the establishment of robust multilayered security system. Financial fraud is committed over cyberspace by cybercriminals using hacking and social engineering approaches to bypass the security. Traditionally fraud detection methods have been employed to the multilayer security system using machine learning and deep learning methods to detect and combat frauds in the financial transactions. However those models face numerous challenges in terms of poor model performance and inaccurate results due to evolving attack strategies and data imbalances (Training of the model with less fraudulent transaction instances). In order to mitigate those challenges, a real time financial transaction processing model has to be designed using deep learning architecture. Thus a new graph convolution network with long short term memory model has been designed with Synthetic Minority Over-sampling Technique (SMOTE). Initially data is preprocessed with missing value imputation using K-NN approach and normalization using min –max normalization. Preprocessed dataset is applied to Synthetic Minority Over-sampling Technique is to eliminate the class imbalance issues occurred in the dataset. It generates the synthetic instances of the fraudulent transaction to minority classes which balances the dataset. Long Short Term Model Network is applied to capture long term dependency of the temporal features of the balanced dataset using gating mechanism along cell state updates and hidden states. Especially dense layer is interfaced to transform the LSTM output into feature vector. Graph Convolution Network transforms the dataset into graph structured data. Graph structured data represented with node and its edge. In thistransaction of the dataset represents the node and transactions relations represent the edge which is obtained using Pearson correlation coefficients. Initially input layer of the graph convolution Network receives the transaction graph and its transaction specific attributes. Further graph convolution layer aggregates information of the neighboring transactions using degree normalized message passing to ensure balanced influence among nodes. The output layer generates updated node embedding which includes features of transaction and aggregated neighbor information. The temporal feature embedding from LSTM model and Relational features embedding from GNN were concentrated to form a unified feature vector. Finally dense layer with activation function and softmax function classifies the transactions with class labels (Fraudulent /Non-Fraudulent) effectively. Experimental analysis of the model is performed using Financial Fraud Detection Dataset on Google colab environment incorporating tensorflow to obtain GPU capabilities. On performance analysis, proposed model attains 96.4% accuracy which found to be better compared to conventional fraud detection approaches.
Introduction
1. Problem Statement
The rise in digital payment adoption has led to a significant increase in online financial fraud. Cybercriminals exploit vulnerabilities using hacking and social engineering, causing large economic losses for financial institutions, despite existing multilayered security systems.
???? 2. Traditional Fraud Detection Methods
Traditional fraud detection models use machine learning (ML) and deep learning (DL) approaches like:
Support Vector Machines (SVM)
Naïve Bayes
Random Forest
Convolutional Neural Networks (CNN)
Recurrent Neural Networks (RNN)
Challenges:
These models struggle with:
Evolving attack strategies
Class imbalances
Low performance on imbalanced datasets
???? 3. Proposed Solution: GCN + LSTM-Based Model
A hybrid deep learning model combining Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) is introduced. It is enhanced by SMOTE to handle class imbalance.
Captures temporal patterns in transaction sequences
Outputs temporal feature embeddings
GCN Network
Converts transaction data into graph form
Learns relational dependencies using neighbors
Outputs relational feature embeddings
Dense Layer + Softmax
Combines temporal and relational features into a unified feature vector
Classifies transactions as Fraudulent or Non-Fraudulent
???? Model Architecture and Hyperparameters
LSTM Settings:
Hyperparameter
Value
Learning Rate
1e-5
Batch Size
11
Epochs
10
Optimizer
Adam
Loss Function
MSE
GCN Settings:
Hyperparameter
Value
Learning Rate
1e-6
Batch Size
15
Epochs
20
Loss Function
Cross Entropy
Activation
ReLU
Dropout Rate
L2 Regularization
???? Experimental Analysis
Conducted in Google Colab
Used Kaggle dataset (60% training, 40% testing)
Grid Search used for hyperparameter optimization
Evaluation based on confusion matrix, precision, recall, and accuracy
Key Result:
The proposed GCN + LSTM model achieved 96.4% detection accuracy
Outperforms traditional models by effectively integrating temporal and relational features
? Advantages of Proposed Model
Addresses class imbalance using SMOTE
Captures both time-dependent patterns (LSTM) and inter-transaction relationships (GCN)
Enhances detection accuracy and reduces false classifications
Scalable to real-time fraud detection applications
Conclusion
In this paper, a graph convolution network with long short term memory model has been designed and implemented along group of preprocessing such as Synthetic Minority Over-sampling Technique (SMOTE) for dataset balancing, K-NN approach for missing value imputation and Min-max approach for normalization. Graph convolution network extracts relational features on processing relational data in different layers of the network and long short term memory model extracts temporal features on processing time series data in different layers of the network. Dense layer were concatenated feature vector in unified form and classified those feature vector into fraudulent and non-fraudulent financial transaction.
Experimental analysis defines efficiency of the configuration towards processing the financial fraud detection dataset and Performance analysis of model reports 96.4 % accuracy which is found to be better compared to accuracy value of the traditional architectures in financial fraud detection. As a future work, accuracy of the fraud detection can be enhanced on employing federated architectures.
References
[1] K. G. Al-Hashedi and P. Magalingam, ‘‘Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019,’’ IEEE Access, vol. 40, 2021
[2] Cherif, A. Badhib, H. Ammar, S. Alshehri, M. Kalkatawi, and A. Imine, ‘‘Credit card fraud detection in the era of disruptive technologies: A systematic review,’’ J. King Saud Univ. Comput. Inf. Sci., vol. 35, no. 1, pp. 145–174, Jan. 2023
[3] J. . Forough and S. Momtazi, ‘‘Ensemble of deep sequential models for credit card fraud detection,’’ Appl. Soft Comput., vol. 99, Feb. 2021,
[4] G. Douzas and F. Bacao, ‘‘Effective data generation for imbalanced learning using conditional generative adversarial networks,’’ Expert Syst. Appl., vol. 91, pp. 464–471, Jan. 2018
[5] 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
[6] 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,
[7] Ienchaji, 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.
[8] S. Nami and M. Shajari, ‘‘Cost-sensitive payment card fraud detection based on dynamic random forest and k-nearest neighbors,’’ Expert Syst. Appl., vol. 110, pp. 381–392, Nov. 2018.
[9] L. Zheng, G. Liu, C. Yan, C. Jiang, M. Zhou, and M. Li, ‘‘Improved TrAdaBoost and its application to transaction fraud detection,’’ IEEE Trans. Computat. Social Syst., vol. 7, no. 5, pp. 1304–1316, Oct. 2020.
[10] E. Esenogho, I. D. Mienye, T. G. Swart, K. Aruleba, and G. Obaido, ‘‘A neural network ensemble with feature engineering for improved credit card fraud detection,’’ IEEE Access, vol. 10, pp. 16400–16407, 2022.
[11] Y. Chen, N. Ashizawa, C. K. Yeo, N. Yanai, and S. Yean, ‘‘Multi-scale selforganizing map assisted deep autoencoding Gaussian mixture model for unsupervised intrusion detection,’’ Knowl.-Based Syst., vol. 224, Jul. 2021,
[12] S. V. Suryanarayana, G. N. Balaji, and G. V. Rao, ‘‘Machine learning approaches for credit card fraud detection,’’ Int. J. Eng. Technol., vol. 7, no. 2, p. 917, Jun. 2018.
[13] Y. Alghofaili, A. Albattah, and M. A. Rassam, ‘‘A financial fraud detection model based on LSTM deep learning technique,’’ J. Appl. Secur. Res., vol. 15, no. 4, pp. 498–516, Oct. 2020
[14] S. A. Ebiaredoh-Mienye, E. Esenogho, and T. G. Swart, ‘‘Artificial neural network technique for improving prediction of credit card default: A stacked sparse autoencoder approach,’’ Int. J. Electr. Comput. Eng. (IJECE), vol. 11, no. 5, p. 4392, Oct. 2021.
[15] M. Mukherjee and M. Khushi, ‘‘SMOTE-ENC: A novel SMOTE-based method to generate synthetic data for nominal and continuous features,’’ Appl. Syst. Innov., vol. 4, no. 1, p. 18, Mar. 2021