Artificial Intelligence Powered Credit Card Fraud Detection System using Random Forest Feature Selection with Convolution Neural Network

Artificial Intelligence Powered Credit Card Fraud Detection System using Random Forest Feature Selection with Convolution Neural Network

Anitha K

Assistant Professor, Department of Computer Science, IT, AI & ML, Srinivasan College of Arts & Science, Perambalur- 621212, Tamilnadu, India,

Sharin N

Assistant Professor, Department of Computer Science, IT, AI & ML, Srinivasan College of Arts & Science, Perambalur- 621212, Tamilnadu, India,

Madhumitha. R

Programmer, Department of Computer Science, IT, AI & ML, Srinivasan College of Arts & Science, Perambalur- 621212, Tamilnadu, India,

Abstract – Credit card fraud detection is increasing, creating financial sector problems due to frauds imperatively growing to create a reliable trust to make communications. Traditional methods in previous research gaps cover the data imbalance and non-feature selection problems, which lead to poor accuracy. To resolve this problem, we propose an Artificial intelligence Credit Card Fraud detection system using Random Forest feature selection with a convolution neural network to effectively identify credit card fraud detection. The dataset is collected from a benchmark repository, including credit feature labels with transaction communication data. The dataset is preprocessed with Synthetic Minority Over-sampling Techniques to address the imbalanced data and normalize the min- max Scaling range to verify the data balancing between the actual values. The optimal random forest algorithm (ORFA) is used to select fraudulent data from a balanced dataset and choose essential data to reduce the variation in target feature values, which is non- dependent. The classification is done with a Convolution neural network (CNN) to effectively identify the fraud data transactions. The proposed system also improves the detection ratal, improving precision, recall, and F1 measures compared to the existing system.

Keywords: Credit Card Fraud Detection, Artificial Intelligence, Random Forest, Convolutional Neural Network, Data Imbalance, SMOTE, Feature Selection, Optimal Random Forest Algorithm, CNN, Financial Security, Transaction Communication Data.

1. INTRODUCTION

   Credit card fraud has become an increasingly pervasive and sophisticated threat, posing significant challenges to the financial sector. The relentless ingenuity of fraudsters necessitates the development of robust and adaptive detection systems to maintain the integrity of financial transactions and preserve customer trust. Traditional fraud detection methods, while foundational, often exhibit limitations in addressing the complexities of modern fraud schemes. A critical research gap lies in effectively handling data imbalance, a common characteristic of fraud datasets where legitimate transactions far outnumber fraudulent ones. Furthermore, the absence of robust feature selection techniques in many existing approaches can lead to suboptimal accuracy, as irrelevant or redundant features can obscure the patterns indicative of fraudulent activity.

   To address these shortcomings, this study proposes an Artificial Intelligence-powered Credit Card Fraud detection system that leverages Random Forest feature selection with a Convolutional Neural Network. By finding the most noticeable elements in the transaction data and then using a deep learning model to categorize transactions as either legitimate or fraudulent accurately, this hybrid technique seeks to improve detection efficiency and accuracy. Because the dataset used in this study comes from a benchmark repository and includes a wide range of credit card attributes and transaction communication data, it offers a realistic and demanding setting for assessing the suggested solution.

   One essential step in reducing the difficulties caused by feature scaling and data imbalance is data preprocessing. The Synthetic Minority Over-Sampling Technique (SMOTE) is used to create fictitious fraudulent transactions to balance the representation of both classes and rectify the dataset’s imbalance. To ensure that no one feature has an undue impact on the model’s learning process, min-max scaling is also used to normalize the feature values within a constant range. Following data preprocessing, the Optimal Random Forest Algorithm (ORFA) selects the most informative features from the balanced dataset. This feature selection step aims to reduce noise and variation in the target feature values, focusing the model’s attention on the most relevant indicators of fraudulent activity. Lastly, a CNN is fed the chosen characteristics to classify them. Accurate detection of fraudulent transactions is made possible by CNN’s ability to recognize intricate patterns and correlations in transaction data. Key performance indicators, including accuracy, recall, and F1-measure, are used to assess the effectiveness of the suggested method. The findings emphasize the efficacy of the suggested AI-powered strategy in preventing credit card fraud by showing a notable

   improvement in detection rate, accuracy, recall, and F1- measure compared to current systems.

2. LITERATURE SURVEY

   State-of-the-art ML and DL methods such as Decision Trees, Random Forests, Support Vector Machines, and Recurrent Neural Networks were used in [1] to improve how fraud is identified. While the approaches performed better, the study still struggled with having more false positives because real-world data is more often biased. The author applied the Synthetic Minority Over-sampling Technique to Adaptive Boosting (AdaBoost) to enhance the presence of underrepresented classes and improve the models sensitivity. Still, although the memory improved, the approach ran into overfitting issues and had lower accuracy with an oversized training set.

   Azharudheen and Vijayalakshmi [3] focus on enhancing big data analysis and protection by integrating privacy- preserving mechanisms directly into an intrusion detection framework for large-scale environments. Working in the context of medical big data and multi-cloud systems, the authors argue that conventional intrusion detection techniques struggle with both scalability and the need to preserve data confidentiality when handling high-volume, high-velocity datasets.

   In a complementary direction, Azharudheen and Vijayalakshmi [4] analyze emerging data protection mechanisms whose aim is to maximize data availability without compromising privacy in pervasive, resource-constrained and cloud-connected environments.

   A team developed a real-time model in [5] incorporating Graph Neural Networks and Autoencoders to strengthen the radio industry’s ability to detect fraud. While showing promise for recording transaction data, the model was expensive to run in real-time because it needed much processing and memory. A sampling method that uses Mahalanobis Distance and Random Forest (RF) was presented in [6] to help classes stand out. Yet, it couldnt keep up with changing transaction behaviors and saw a fall in accuracy whenever outliers existed.

   The author used a Random Under Sampling approach (RUS) and Two-Stage Thresholding to reduce false positives in the ensemble learning process. It becomes a problem in this model when some legitimate samples are lost because they are underrepresented. In [8], a more advanced system named Improved Variational Autoencoder with a Generative Adversarial Network (IVAE-GAN) was introduced to create realistic examples of fraud. Still, giving more weight to minors helped, yet unstable learning and collapsing modes kept it from being very useful.

3. PROPOSED METHODOLOGIES FOR AN AI-POWERED CREDIT CARD FRAUD DETECTION SYSTEM

   Credit card fraud poses a significant and escalating threat to the financial sector, eroding trust and causing substantial monetary losses. Developing robust and reliable fraud detection systems is paramount, particularly in the face of increasingly sophisticated fraudulent activities. Traditional methods often fall short due to inherent challenges such as data imbalance and suboptimal feature selection, resulting in poor accuracy and limited effectiveness. This essay proposes a novel, AI-powered credit card fraud detection system designed to overcome these limitations, leveraging Random Forest feature selection in conjunction with a Convolutional Neural Network to achieve enhanced accuracy and performance.

   Communication

   Trans Logs

   Credit card Logs

   Feature Labels and Margins

   SMOTE sampling

   Adaptive Random Forest (ARF)

   Identified credit

   logs

   Classification -CNN Unit

   Figure 1: Work process for ARF-CNN

   The proposed methodology comprises several key stages: data collection and preprocessing, feature selection using an Optimized Random Forest Algorithm (ORFA), and fraud detection via a CNN. Figure 1 shows the Work process for ARF-CNN. Each stage is carefully designed to address the shortcomings of traditional methods and contribute to a more robust and reliable system.

   1. Data Collection and Preprocessing

      The foundation of any successful fraud detection system lies in the quality and comprehensiveness of the data used for training and evaluation. The proposed system will utilize a benchmark dataset containing credit feature labels alongside transaction communication data. This dataset, representative of real-world credit card transactions, will provide a rich source of information for identifying fraudulent patterns.

      Addressing the inherent data imbalance in fraud detection datasets where legal transactions greatly outweigh fraudulent ones is crucial to data preprocessing. The model may be skewed towards forecasting valid transactions due to this imbalance, which might result in low detection rates for fraudulent activity. The Synthetic Minority Over-Sampling Technique (SMOTE) will be used to lessen this problem. To adequately balance the dataset and avoid bias, SMOTE creates synthetic samples of the minority class (fraudulent transactions) based on the minority class samples that are already available. The issue that fraudulent cases make up a tiny percentage of legitimate ones is addressed by applying the Synthetic Minority Over-Sampling Technique to the Credit Card Fraud Detection dataset. During the preparation phase in the proposed method, synthetic examples of the minority (fraudulent) group are created using SMOTE by interpolating between the given minority class records. With this approach, we identify fraudulent transactions without repeating data, which helps the model fit the data appropriately. When using SMOTE, the system can more accurately find differences between each class and better identify rare types of fraud without losing the ability to understand general transactions. The preprocessing method is significant since it stops the credit card fraud detection framework from being affected by the majority of the data and improves its accuracy and reliability.

      A significant difficulty in this dataset is that there are far more normal transactions (the primary class) than fraudulent ones (the minority class). Our approach solves this by using the SMOTE method for data preparation. If indicates a fraudulent transaction, represents a sample from . In SMOTE, one of the –

      nearest neighbors () of the marked sample is selected using a measure such as Euclidean distance. Subsequently, a synthetic sample x_new is formed by connecting and with equation 1.

      = + . ( ) (1)

      Where comes from a uniform distribution on the range (0,1), as a result of this method, new data is created to help diversify the minority class, keeping the original data clean. To establish the required number of synthetic samples , the

      method examines the number of majority class instances and minority class instances . The number of synthetic cases required for a balancing ratio is calculated as

      = . (2)

      With this approach, the number of students in each class is better spread out to fulfill the desired goals (like a 1:1 ratio).

      Next, dataset is formed by joining the original samples and synthesized images using equation 3.

      = (3)

      This case uses to mean the original legitimate ones, to mean the original fake ones, and to tell the newer fake transactions made by the attacker. Improvements to the model’s ability to find similarities in the data of both kinds of behavior allow it to catch both regular fraud cases and the more unusual ones.

      Furthermore, data normalization is crucial for ensuring that all features contribute equally to the learning process and preventing features with larger scales from dominating the model. Min-Max scaling will normalize the data within a defined range, typically between 0 and 1, ensuring that all feature values are scaled proportionally. This step ensures data consistency and optimizes the performance of subsequent algorithms.

   2. Feature Selection using Optimized Random Forest Algorithm (ORFA):

      Feature selection plays a pivotal role in improving the accuracy and efficiency of the fraud detection system. Selecting the most relevant features reduces noise, minimizes redundancy, and enhances the model’s ability to generalize to unseen data. Traditional feature selection methods often fail to address the specific characteristics of fraud detection datasets adequately. Deep learning approaches to credit card fraud detection can enhance the effectiveness of fraud detection systems when supplemented with feature selection measures like Random Forest. This implies that the algorithm uses a Random Forest as a feature selection method. The Random Forest feature selection method allows the model to retain only the most meaningful features, reducing dimensionality, removing noise, and improving generalization.

      = {(, ), = 1,2, } (4)

      where S is the original dataset, is the feature vector for the transaction, is the corresponding label, and N is the total number of transactions.

      = {(, ), = 1,2, . , } (5)

      Where C is the selected feature subset after applying random forest, is a reduced feature vector containing only the c most important features, and C is the number of features retained after selection.

      = {(, ), = 1,2, } (6)

      D can represent another derived transformed feature set, possibly after some additional dimensionality reduction transformation, and d is the number of transformed features.

      (, ) = () (12)

      2

      (7)

      Where (, ) are the margin score for query q with respect to the variable, a () is the average importance of the feature, and E* is the expected value of a baseline comparison measure.

      (, ) = .

      ||||×||||

      (8)

      where (, ) is the cosine similarity between two feature vectors a and b, and

      .

      ||||×||||

      are the dot products of the vectors, magnitudes of vectors a and b. These are used to measure similarity between two data points’ features, often to identify redundant correlated features.

      The proposed system employs an Optimized Random Forest Algorithm for feature selection. Random Forest, an ensemble learning method, is inherently robust to outliers and capable of handling high-dimensional data. The “Optimized” component enhances the standard Random Forest algorithm by incorporating a customized optimization strategy. This strategy aims to identify and selectfeatures highly predictive of fraudulent transactions while minimizing the inclusion of non-dependent or irrelevant features. The ORFA works by iteratively evaluating subsets of features based on their contribution to the model’s predictive performance, optimizing for a balance between accuracy and feature set size. By focusing on the most impactful features, the ORFA reduces the dimensionality of the data, leading to improved model performance and reduced computational complexity.

   3. Fraud Detection via Convolutional Neural Network (CNN):

      Convolutional Neural Networks, typically associated with image processing, offer a powerful approach to identifying patterns and anomalies in sequential data, making them well-suited for fraud detection. In this context, transaction data can be structured as a sequence of features, allowing CNN to learn temporal dependencies and intricate patterns indicative of fraudulent activity. Convolutional Neural Networks have typically been used for image-processing tasks. Still, they are very effective in classification problems involving structured data (transaction records) and have been used in fraud detection. In the case of fraud detection, CNNs can be used to teach an enumerable feature vector as a one-dimensional “image” or sequence, which allows the CNN to learn spatial or temporal patterns in the data automatically.

      ( ), = [ + , + ]. [, ] (9)

      Where ( ), input feature matrix, and [, ] are filter weights at position (q,n), [ + , + ] input values overlapping, this equation represents the core convolution step.

      () = max(0, ) (10)

      Where () is output after applying the ReLU activation, and j are input values from the previous layer, non-linearity means the network learns complex patterns by zeroing out negative values.

      , = max (|, |) (11)

      Where , are the result of the max pooling operation at position (y,z), and |, | is a local region of the input data over the maximum taken, pooling in down sampling and reducing dimensionality and making features more robust to small changes in the input

      = () + (12)

      Where is the index after transformation, and () is mapping functions that might represent layer depth, this equation relates to tracking feature positions data reshaped between layers.

      The CNN architecture will be designed to capture both local and global patterns in the transaction data. Convolutional layers will extract relevant features by applying filters that slide across the input sequence, identifying fraud-related characteristic patterns. Pooling layers will down-sample the feature maps to decrease the dimensionality of the data and increase the model’s resilience to changes in the input data. After extracting characteristics, fully linked layers will combine to create a final classification that will mark each transaction as authentic or fraudulent.

      The CNN will be trained using the preprocessed and feature-selected data. During training, the network will learn to identify patterns and anomalies indicative of fraudulent transactions. A held-out validation set will be used to regularly assess the model’s performance to ensure it avoids overfitting and generalizes effectively to new data.

4. Results and Evaluation Metrics:

   The proposed AI-powered credit card fraud detection system is expected to improve fraud detection rates compared to traditional methods significantly. The combination of SMOTE for data balancing, ORFA for feature selection, and CNN for classification addresses the key challenges associated with fraud detection, including data imbalance, high dimensionality, and complex patterns.

   1. Comparison performance results

      Table 1 Comparison results

      Model	Precision	Recall	FI-Measure	AUC-ROC
      Logistic Regression	0.84	0.76	0.80	0.89
      Support Vector Machine	0.81	0.70	0.75	0.86
      Decision Tree	0.86	0.78	0.82	0.90
      Convolutional Neural Network	0.91	0.88	0.90	0.95

      Table 1 compares different models for credit card fraud detection in terms of their performance metrics. The CNN model is the best performer, with a precision of 0.91, recall of 0.88, F1-measure of 0.90, and AUC- ROC of 0.95. Logistic Regression and SVM perform moderately, while the Decision Tree performs slightly better.

      Figure 2: Illustrates Precision Performance

      In Figure 2, the x-axis shows the precision scores, while the y-axis lists the models in ascending order of performance. Each grey dot represents an individual evaluation or run, and the orange line represents the trend or average performance across the models. The precision values indicate that the CNN had the highest precision score of 0.91, followed by the Decision Tree, with a score of 0.86. The values for Logistic Regression and SVM were 0.84 and 0.81, respectively.

      Figure 3: Illustrates Recall Performance

      Figure 3 illustrates the x-axis of recall values, with the models listed along the y-axis in ascending order of recall performance. As for the figure, it displays a scatter plot with multiple blue dots representing individual performance points, with a single trend line depicting the overall pattern. The CNN model has the best recall at the point of 0.88, indicating it was the most successful in correctly identifying all relevant instances for the outcome variable. The decision tree followed with a recall of 0.78, logistic regression at 0.76, and the SVM had the lowest overall recall at 0.70.

      Figure 4: Illustrates FI-Measure Performance

      Figure 4 presents the performance of the four classification models, LR, SVM, DT, and CNN, based on the F1-Measure. The x-axis represents F1-measure values, and the y-axis identifies the models in order, starting with the lowest performance. Each blue dot represents an observation, and the trend line shows an overall improvement through the models. CNN with the F1-Measure of

      0.90 has the highest overall modeling performance, showing the best balance between precision and recall, followed by Decision

      Tree (0.82), LR (0.80), and SVM with the lower score of (0.75).

      Figure 5: Illustrates FI-Measure Performance

      Figure 5 shows the AUC-ROC performance from SVM, DT, and CNN from four classification models. The x-axis shows AUC-ROC scores, and the y-axis lists models in ascending order based on performance. CNN has the highest AUC-ROC score of 0.95 since it has the highest ability to separate the classes. The DT model is second with a score of 0.90, the SVM score is 0.86, and LR scored 0.89, respectively.

5. CONCLUSION:

The proposed AI-powered credit card fraud detection system represents a significant advancement in the fight against financial fraud. By combining state-of-the-art techniques such as SMOTE, ORFA, and CNN, the system addresses the inherent challenges of data imbalance, feature selection, and pattern recognition. The expected outcomes include improved fraud detection rates, reduced false positives, and enhanced overall performance compared to traditional methods. The rigorous evaluation framework will ensure that the system meets the highest standards of accuracy and reliability, providing a robust and effective solution for protecting the financial sector from the ever-increasing threat of credit card fraud. The successful implementation of this system will not only mitigate financial losses but also enhance trust and confidence in the credit card ecosystem.

REFERENCE

1. F. K. Alarfaj, I. Malik, H. U. Khan, N. Almusallam, M. Ramzan and M. Ahmed, “Credit Card Fraud Detection Using State-of-the-Art Machine Learning and Deep Learning Algorithms,” in IEEE Access, vol. 10, pp. 39700-39715, 2022, doi: 10.1109/ACCESS.2022.3166891.
2. E. Ileberi, Y. Sun and Z. Wang, “Performance Evaluation of Machine Learning Methods for Credit Card Fraud Detection Using SMOTE and AdaBoost,” in IEEE Access, vol. 9, pp. 165286-165294, 2021, doi: 10.1109/ACCESS.2021.3134330.
3. A. Mohamed Azharudheen and Dr.V. Vijayalakshmi, Improvement of data analysis and protection using novel privacy-preserving methods for big data application The Scientific Temper Vol. 15, no. 2, pp. 2181-2189, 2024.
4. A. Mohamed Azharudheen and Dr.V. Vijayalakshmi, Analyze the New Data Protection Mechanism to Maximize Data Availability without Having Compromise Data Privacy Educational Administration: Theory and Practice, Vol.30. No.5, pp. 3911-3922, 2024.
5. I. D. Mienye and Y. Sun, “A Deep Learning Ensemble With Data Resampling for Credit Card Fraud Detection,” in IEEE Access, vol. 11, pp. 30628-30638, 2023, doi: 10.1109/ACCESS.2023.3262020.
6. A. A. Taha and S. J. Malebary, “An Intelligent Approach to Credit Card Fraud Detection Using an Optimized Light Gradient Boosting Machine,” in IEEE Access, vol. 8, pp. 25579-25587, 2020, doi: 10.1109/ACCESS.2020.2971354.
7. F. Khaled Alarfaj and S. Shahzadi, “Enhancing Fraud Detection in Banking With Deep Learning: Graph Neural Networks and Autoencoders for Real-Time Credit Card Fraud Prevention,” in IEEE Access, vol. 13, pp. 20633-20646, 2025, doi: 10.1109/ACCESS.2024.3466288.
8. Z. Xie and X. Huang, “A Credit Card Fraud Detection Method Based on Mahalanobis Distance Hybrid Sampling and Random Forest Algorithm,” in IEEE Access, vol. 12, pp. 162788-162798, 2024, doi: 10.1109/ACCESS.2024.3421316.
9. I. Almubark, “Advanced Credit Card Fraud Detection: An Ensemble Learning Using Random Under Sampling and Two-Stage Thresholding,” in IEEE Access, vol. 12, pp. 192079-192089, 2024, doi: 10.1109/ACCESS.2024.3519335.
10. Y. Ding, W. Kang, J. Feng, B. Peng, and A. Yang, “Credit Card Fraud Detection Based on Improved Variational Autoencoder Generative Adversarial Network,” in IEEE Access, vol. 11, pp. 83680-83691, 2023, doi: 10.1109/ACCESS.2023.3302339.
11. A. Mniai, M. Tarik and K. Jebari, “A Novel Framework for Credit Card Fraud Detection,” in IEEE Access, vol. 11, pp. 112776-112786, 2023, doi: 10.1109/ACCESS.2023.3323842.
12. H. S. Alsagri, “Hybrid Machine Learning-Based Multi-Stage Framework for Detection of Credit Card Anomalies and Fraud,” in IEEE Access, vol. 13, pp. 77039-77048, 2025, doi: 10.1109/ACCESS.2025.3565612.